2% Equity, $90K Base: The YC Startup Betting Big on Two Engineers

Two Roles, One Signal

A three-person team out of Y Combinator's S25 batch is hiring two engineers to build the infrastructure that lets autonomous agents trade real money. The roles are distinct. The signal they send together is not: Conviction needs someone who can ship production-grade agentic systems today, and someone who can make those systems financially literate tomorrow.

Conviction lists two openings in San Francisco, both open to visa sponsorship. The Founding Engineer role is the first engineer through the door, working directly with CEO Aria Vikram and CTO Tony Yamin on multi-agent systems, backend infrastructure, and the financial data pipelines that power research and live trading. The Quantitative Engineer sits alongside it. The title alone tells the split: one seat for the builder who makes agents reliable at scale, one for the engineer who makes them smart about markets.

The Founding Engineer description reads like a checklist for a senior infrastructure generalist. Required: AWS primitives (ECS, ALBs, S3, RDS), deep PostgreSQL fundamentals including schema design and connection pooling, Docker and GitHub Actions CI/CD, and — crucially — experience building production-grade agentic systems. The posting also calls out concurrency, multithreading, and asynchronous programming — table stakes for a system that ingests market data, news sentiment, SEC filings, earnings, price action, volume, and volatility simultaneously without falling over.

Financial domain knowledge is a plus, not a requirement. Hedge fund, prop trading, asset management, fintech, or quant research experience "is a plus, but not required." The phrasing is deliberate. Conviction's product uses that same data suite, plus indicators, to build informed strategies and reduces overfitting by testing across regimes. But the founders are not hiring a quant to hand-tune models. They are hiring an engineer to build the scaffolding that lets agents reason over that data reliably. The Quantitative Engineer role exists because that scaffolding needs a counterpart who speaks the language of alpha, risk, and execution.

The pair of roles reveals a company at a specific inflection point. With three people total — Y Combinator's figures put the team at three — (two founders and one existing team member), Conviction has moved past the prototype phase into the production phase. The Founding Engineer owns the path from idea to working product, setting technical and cultural foundations. The Quantitative Engineer ensures the product's decisions are grounded in financial reality. Together they represent the two hardest problems in agentic finance: reliability and domain competence. Conviction is betting it can hire for both simultaneously.

What the Resume Screen Actually Catches

That band — Zero G Talent's board data shows $90K–$210K, and Y Combinator's posting lists 0.5%–2% equity — acts as a coarse filter. Conviction is not competing on cash with hyperscalers or quant funds that top $300K for similar titles. It competes on equity upside, autonomy, and problem density. The resume screen implicitly down-weights candidates whose primary lever is compensation maximization and up-weights those who have previously taken below-market cash for asymmetric upside: prior founding-engineer stints, early employee at an acquired startup, or a quant role where they left carry on the table to build something they controlled.

A quantitative engineer hiring at this stage signals a need for production-grade modeling chops, not research prototypes. The title "quantitative" over "ML" or "AI" suggests the team expects candidates to own the full stack from data ingestion through execution latency, with heavy emphasis on statistical rigor and risk controls. A founding engineer role, by contrast, is a breadth filter: it selects for engineers who have shipped systems end-to-end in ambiguous, high-velocity environments, often without dedicated infra or product support. The overlap is narrow. Candidates who clear both filters tend to show three things on paper: a track record of moving models or quantitative strategies into live trading or production environments where P&L attaches directly to code decisions; fluency in the Python/C++/Rust toolchain that actually runs at microsecond-to-millisecond loops, not just notebook experimentation; evidence of owning a meaningful slice of a system (data pipeline, execution engine, risk layer) rather than contributing tickets to someone else's architecture.

Broader research on hiring screens (Uber's 3.5 million rejections since 2017, EEOC guidance on individualized assessment, California's Fair Chance Act timing rules, Illinois' new disparate-impact framework) shows that legal screening floors are rising and standardizing. Conviction's screen operates well above those floors. Its Y Combinator backing and agentic-trading focus place it in a talent tier where resume review is less about disqualification and more about signal extraction: does this person understand the difference between a model that backtests and a system that survives a flash crash? Have they debugged a production incident where the cost was measured in dollars per second? Can they articulate the tradeoff between model complexity and operational fragility without reaching for buzzwords?

No research shows Conviction publishing a "we don't hire from FAANG" or "we require open-source contributions" rule. The two roles reveal a company building a thin, high-leverage engineering team where every hire must expand the system's surface area. The resume that gets a call proves the candidate has already operated at that leverage, whether in a prop shop, a high-frequency market-making desk, a crypto market-maker, or a prior founding-engineer seat where the product was the infrastructure. Everything else is noise.

How the Interview Loop Works

Conviction has not published its interview loop. Board data covers only the two open roles and the founders. What follows maps the pattern early-stage, YC-backed AI startups building agentic systems run, to the specific demands of a trading platform that must execute reliably with real money on the line.

Screen: Founder or Technical Lead Call (30–45 minutes)

The first live conversation is almost always with a founder (Yamin or Vikram) or a senior engineer they trust. The goal is not to test syntax but to verify that the resume reflects actual ownership of complex systems. Expect questions about the hardest scaling or correctness problem you have solved end-to-end, how you instrumented it, and what you would change if you rebuilt it today. For the Quantitative Engineer role, the discussion tilts toward statistical rigor: walk me through a model that failed in production, how you detected the drift, and what guardrails you added. For the Founding Engineer role, the probe shifts to product judgment: describe a time you chose a boring technology over a flashy one because the risk profile demanded it. Candidates who speak in abstractions ("we built a robust pipeline") get cut. Candidates who cite concrete metrics ("latency dropped from 120 ms to 18 ms at p99 after we rewrote the batcher in Rust") advance.

Deep-Dive: System Design + Agentic Architecture (60–90 minutes)

This round filters for production-grade agentic experience. The interviewer, often a founder paired with a senior engineer, presents a scenario drawn from Conviction's actual roadmap: an autonomous agent that must ingest market data, reason over a strategy, and place orders with sub-second latency while maintaining an auditable decision log. Evaluation runs on three axes. First, decomposition: can you break the problem into bounded services (ingestion, reasoning, execution, observability) with clear contracts? Second, failure modeling: do you design for partial degradation — circuit breakers, idempotency keys, replayable event logs — or do you assume the happy path? Third, tooling fluency: the stack is Python-heavy with Rust for hot paths, so expect to discuss async patterns, memory management, and why you would or would not reach for a framework like LangGraph versus a hand-rolled state machine. A candidate who cannot articulate the trade-off between LLM-as-judge evaluation and deterministic unit tests for agent behavior will not pass.

Live Coding: Correctness Under Constraints (60 minutes)

Unlike big-tech loops that optimize for algorithmic cleverness, this session optimizes for production hygiene. You write code in a shared editor — no whiteboard — typically a refactoring exercise: here is a 200-line module that processes order-book snapshots; it has a race condition, a memory leak, and no structured logging. Fix it. The interviewer watches for: do you write a failing test first? Do you introduce types (mypy/pydantic) to make the contract explicit? Do you ask what the SLA is before choosing a data structure? The Quantitative Engineer track adds a numerical stability component: implement a rolling z-score that does not blow up on zero-variance windows. The Founding Engineer track adds a deployment question: how would you canary this change to 1% of traffic with automatic rollback on error-rate spike?

Pair Programming / Debugging Session (60–90 minutes)

A newer stage adopted by several YC AI companies in the past 18 months: you join an engineer on a real ticket from the backlog, perhaps adding a new tool to the agent's repertoire or hardening the replay subsystem. This is not a performance; it is a collaboration signal. The team watches how you ask for context, how you handle an unfamiliar codebase, and whether you leave the code better than you found it (tests, docs, type hints). For a two-person founding team hiring their first engineers, this round carries disproportionate weight — it is the closest proxy to "what is it like to ship with this person on day 30?"

There is no right answer — only consistency with the culture they are building.

The loop is short by design (four to five conversations over two weeks) because Conviction cannot afford a three-month process and candidates at this level typically have competing offers. Every stage is eliminative; there is no "culture fit" round that exists to be polite. If you clear the system-design and live-coding rounds, the final conversation confirms. It does not gate.

Your Portfolio Is the Proof

In 2026, your GitHub portfolio is the single most predictive signal of agent engineering ability. That assessment, from a Techademy analysis of hiring patterns across agentic AI teams, reflects a shift Conviction's screen embodies: a resume listing "LLM experience" no longer separates candidates. The company's agentic trading platform demands engineers who have shipped systems that reason, act, and recover under real constraints — not notebooks that work on clean data.

The research draws a bright line. "A project without evaluation is a demo. A project with evaluation is engineering." Production-grade agentic systems require a stack that goes well beyond prompt chains: security, authentication, authorization, monitoring, reliability, scaling, persistent storage, rate limiting, failure recovery, evaluation, and cost management. The Agentic AI Portfolio repository maintained by Sciddhanto Sinha explicitly distinguishes "Working Prototype" from "Production Extension" along these dimensions. Candidates who want to pass Conviction's screen should be ready to walk through how they implemented at least three of those capabilities in a live system.

Progression matters more than breadth. Techademy's recommended sequence (Customer Support Agent, Deep Research Agent, Coding Assistant or Lead Gen Agent, then optionally a Multi-Agent Debate system) mirrors the complexity ladder early-stage teams actually climb. Spacing projects six to eight weeks apart gives time to ship each one to a high standard rather than accumulating half-finished proofs of concept. The strongest deep research agents seen as portfolio pieces shared one trait: they were honest about limitations. If your multi-agent debate system does not outperform a single agent on at least one well-defined metric, do not ship it.

For a trading platform, the relevant metrics are latency, determinism under load, and auditability. Conviction's Quantitative Engineer and Founding Engineer roles will evaluate whether you have built agents that can execute multi-step financial workflows — data ingestion, strategy formulation, risk checks, allocation, reporting — with observable decision traces. The aralla-ai project roadmap maps this domain: Foundations & Data Ingestion, Core Agents (Strategy, Risk, Allocator, Reporting), Unstructured Intelligence (News Sentiment), Observability & Guardrails, Optimization & Streamlit App, Documentation & Deployment. A portfolio that shows this progression, with live endpoints and eval dashboards, signals the exact profile Conviction needs.

Production evidence also means showing you have wrestled with the unglamorous 80%: data engineering, stakeholder alignment, governance, and workflow integration. MIT Sloan research on agentic deployments found those tasks consumed the majority of effort in a clinical adverse-event detection system. Candidates who can describe a concrete incident — a schema drift that broke an agent's tool call, a guardrail that prevented a runaway position, a cost spike from unbounded token usage — and how they instrumented detection and recovery will stand out.

Finally, demonstrate you understand the evaluation loop. JPMorgan Chase's LAW system achieves 92.9% accuracy across legal queries because it measures specialized agents against ground truth continuously. Rakuten cut feature time to market by 79% by running five tasks in parallel with Anthropic's Claude Code, but only after building the observability to verify correctness. Coinbase's agent-first model delivered a 90% reduction from idea to production because they redesigned sprint planning around measurable agent outcomes. Conviction's screen will probe whether you treat evaluation as a first-class engineering discipline, not an afterthought.

Culture Is the Daily Interaction

Conviction's YC backing adds a fourth filter: speed of iteration. YC partners push founders to ship weekly, measure weekly, pivot weekly. An engineer who optimizes for elegance over learning velocity will frustrate the loop. The interview process, whatever its formal stages, probes for evidence of past projects where the candidate shipped fast, measured honestly, and changed course without drama.

The salary band itself carries cultural information. A $90K floor in San Francisco is below market for senior AI engineers, but the $210K ceiling reaches well above typical early-stage equity-heavy offers. That spread suggests Conviction is willing to pay for experience but expects candidates to bet on upside. Someone who negotiates hard on base and ignores equity signals misalignment with the founder mindset.

Early-stage AI startups tend to filter for three non-technical traits that rarely appear in job descriptions. First, comfort with ambiguity. An agentic trading platform has no settled architecture; the product will be defined by the engineers who build it. Candidates who need a spec before they write code will stall. Second, ownership density. With two engineers, there is no "platform team" to offload infrastructure to, no "ML ops" squad to hand models off to. Each hire must treat the whole stack (data pipelines, model serving, risk controls, compliance hooks) as their problem. Third, communication cadence. In a pair, every design decision is a conversation. Engineers who disappear for a week and return with a finished module create merge conflicts that are social, not technical.

What the research does not show, and what no external source can reliably confirm, is whether Conviction uses a structured values interview, a trial project, a reference deep-dive, or a founder dinner. The absence of public detail is itself a data point: the cultural filter is informal, high-bandwidth, and concentrated in the founders' hands. For a candidate, the "culture" conversation starts the moment a founder reads the resume and does not end until an offer is signed.

In a two-engineer company, culture is not a program. It is the daily interaction between two people who decide, together, what the product becomes.

Candidates who treat the process as a series of hoops (technical screen, system design, behavioral) will miss the real evaluation. The filter is continuous. Every email response, every question about the roadmap, every follow-up on a technical detail is a data point on whether this person can co-lead a zero-to-one agentic system with the other hire. The technical bar gets you in the room. The cultural bar decides whether you stay.

Ten Ways to Get Rejected

Conviction's screen inherits YC's filter: specificity wins, vagueness loses. The accelerator's 1.5–2% acceptance rate across 20,000-plus applications per batch means partners have seen every pattern. They reject the same flaws repeatedly. Conviction applies that same lens to its two roles. Candidates who treat the application like a pitch deck get filtered fast.

The root cause is consistent. YC partners are not looking for the most ambitious application. They are looking for the most credible one. Credibility comes from specificity. Vagueness, even well-written, confident vagueness, reads as a signal that the founder does not yet have a real grasp on what they are building. For engineers, that translates directly: "I built scalable AI systems" fails. "I shipped a RAG pipeline that cut hallucination rate from 18% to 3% on 50,000 legal documents using Llama-3-70B with custom evals" passes. The first could be fifty people. The second is one person with receipts.

Mistake one: writing a vision statement instead of a role description. Candidates lead with "passionate about agentic trading" or "want to revolutionize finance." Conviction's screen looks for what you have shipped. The YC Library flags "platform," "ecosystem," "solution," "disruptive," "paradigm shift," "world-class," "passionate," "leverage" as a verb, "synergies," "holistic," and any phrase borrowed from Paul Graham essays. Strip them. Replace every adjective with a mechanism. If a word does not change what you built, cut it.

Mistake two: describing traction vaguely. "Strong interest from hedge funds" means nothing. "Three prop-shop PMs ran our alpha signal on paper for six weeks; two asked for API access" means everything. Every traction claim needs a number. Every number needs a unit. Past tense only. Future tense kills credibility. If you have no production deployment, say so plainly and describe the test you are running this week.

Mistake three: citing market research instead of user evidence. YC wants founder-generated evidence: conversations you had, things users told you, behavior you observed, money someone paid you. For an engineer applying to Conviction, that means naming the specific problem you solved for a specific user, not the TAM. "Two NYC dental practices" beats "early adopters in healthcare." The same logic applies: name the trading desk, the strategy, the latency constraint you hit.

Mistake four: claiming no competition. Saying "no competitors" is a near-instant rejection signal per the YC Library. There is always a competitor, even if it is the spreadsheet your customers use today. For agentic trading, the competitor is the existing Python stack the quant team maintains. Acknowledge it. Explain your differentiation specifically.

Mistake five: confusing interest with demand. YC looks for demand signals: money exchanged, repeated usage, referrals, people asking when they can have access. A GitHub star is not demand. A PR merged into a production repo is. A coffee chat with a PM is not demand. That PM pushing your code to staging is.

Mistake six: describing your background without tying it to this problem. Every background detail you include should connect to why it makes you specifically well-suited to build this specific product. "Four years at Stripe building the dispute-handling pipeline, shipped nine production payments systems" beats "experienced backend engineer." The former proves you know the class of problem Conviction faces: latency, correctness, auditability. The latter proves you have a LinkedIn profile.

Mistake seven: letting AI write the substance. Using AI to help structure or edit your application is not inherently a mistake, but letting AI write the substance for you almost always is. AI-generated answers tend to be comprehensive, balanced, and generic, the exact opposite of what a strong YC application needs. The application should sound like you: specific, direct, and shaped by your actual experience with the problem. Conviction's founders will read your words. They know the difference.

Mistake eight: inconsistency between written answers and the video or live screen. Before submitting, watch your video and read your application in the same sitting. Make sure they tell the same story. YC partners stop reading when the narrative fractures. Conviction's loop will do the same.

Mistake nine: submitting without having someone external read it. The single cheapest improvement you can make is having someone unfamiliar with your work read every answer and tell you what questions each answer raises. If a partner has to guess what you mean in any field, you have rewriting to do before the deadline.

Mistake ten: applying too early, or not applying again after rejection. About 60% of the last YC batch had previously applied. Reapplying with refined metrics, sharper answers, and new evidence works. The same holds for Conviction: a "no" now with a follow-up in three months showing a shipped feature, a benchmark result, or a production incident you resolved changes the signal.

The pattern across all ten: evidence over assertion. Conviction's two open roles map to those same dual challenges. The screen is built to catch engineers who have already solved them.

Working in AI? Zero G Talent tracks the openings: see every open Conviction role, browse AI jobs, the companies hiring, and the people building the field.