Steven NakhwalFounder, Operon Labs
New York · Quantitative finance · Production AI

Finance runs on systems that can't fail. I build the AI that runs them.

I spent my early career pricing derivatives at BNY Mellon and building in-house automation at Silver Slate Group. Operon Labs is what that turned into: a small studio that builds AI for finance and revenue teams, with the standards of a trading desk and the suspicion of a risk one.

CurrentlyOnboarding VisionGuard's first agency client and finishing the Netstay forecasting engine.
BaseNew York City
FocusFinance & RevOps automation, ML forecasting, agentic workflows
PreviouslyQuant Analyst, BNY Mellon · Automation engineering, Silver Slate Group
EducationDual BBA, Baruch Zicklin
Operon Labs

An AI automation studio for finance and revenue operations.

I run Operon Labs with a team of three to four engineers with extensive backgrounds in quantitative finance, risk and production machine learning. We take the work that sits between a finance team's systems of record (billing, CRM, ERP, ledgers) and the people who have to reconcile them, and we build software that does it reliably.

Engagements are scoped, built and handed over with monitoring, audit trails and documentation. We don't ship demos; we ship systems that keep running after we leave. Every agent we build goes through the same discipline: locked-down data access, an evaluation suite it has to pass, and a red-team pass before it touches anything real.

operonlab.io →

  • RevenueRevenue & AR automation

    Invoice chasing, collections workflows and reconciliation agents tied to the system of record.

  • RiskRisk & forecasting systems

    Pricing, forecasting and monitoring models built with walk-forward validation, not notebook screenshots.

  • AgentsAgentic back-office workflows

    Chat and voice agents, document pipelines and compliance checks with append-only audit logs.

  • BuildProduct engineering

    Full-stack builds for products that need an ML core: FastAPI, Postgres, vector search, React.

What makes an agent strong

Wiring a model to a tool takes an afternoon. Getting a finance team to trust what comes out takes everything else on this list.

Data security

Schema-per-tenant isolation, encryption at rest and in transit, least-privilege tool access scoped to the task. The agent sees the rows it needs and nothing else, and every read is logged.

In practice AES-256-GCM · per-tenant Postgres schemas · scoped service credentials · no prompt ever carries a secret

Evaluation systems

Before an agent ships it has a benchmark: real historical cases with known-correct outcomes, scored on accuracy, cost and latency. Every prompt or model change re-runs the suite. No eval, no deploy.

In practice golden datasets from client history · regression gates in CI · LLM-as-judge with human-labelled calibration sets

Red teaming

We attack our own agents before anyone else does: prompt injection through documents and emails, tool-call abuse, data exfiltration attempts, adversarial inputs at the edges of the policy. Findings become test cases.

In practice injection corpora · tool-permission fuzzing · exfiltration probes · abuse cases written into the eval suite

Guardrails & audit

Hard limits the model can't talk its way past: spending caps, allow-listed actions, human approval on anything irreversible, and an append-only signed log of every decision so a dispute six months later has a paper trail.

In practice HMAC-signed audit logs · kill switches · approval queues · policy checks outside the model

Observability

Traces for every run, drift monitors on inputs and outputs, and alerting tied to business metrics, not just uptime. When the agent's quality slips, we know before the client does.

In practice per-run tracing · output drift detection · cost and latency budgets · on-call ownership after handover

Walk-forward validation

Forecasting and trading models are judged only on data they never saw, in rolling out-of-sample folds. In-sample results are not reported, to clients or to ourselves.

In practice rolling train/test folds · calibration checks · per-property or per-desk recalibration

Production systems

Netstay

BuildingHospitality · Forecasting

AI cancellation forecasting for hotels. Netstay scores every future reservation on the books, testing dozens of variables at once (channel, lead time, rate, length of stay, guest history, seasonality, local events) to flag the bookings most likely to cancel and when. Revenue teams see tonight's and next month's real expected occupancy instead of the reservation count, so they can overbook with confidence, reprice early and stay full. Every empty room a cancellation leaves behind is cost the hotel never recovers; Netstay exists to take that number toward zero.

Stack Gradient-boosted and sequence models over booking, rate and lead-time features · per-property calibration · walk-forward backtests · standalone product codebase

Netstay · Property 0412 · Next 30 daysnetstay.app/p/0412/forecast

Occupancy: on the books vs expected after cancellationsrooms / night

180150120Oct 16Oct 26Nov 5Nov 14 Oct 29 · 153 expected (171 OTB)
On the booksExpected80% interval

Highest cancel risknext 14 d

BookingArriveP(cancel)
BK-77124 · OTAOct 220.81
BK-77003 · OTAOct 190.74
BK-76880 · DirectOct 250.58
BK-77210 · CorpOct 240.52
BK-76991 · OTAOct 200.47
Model v0.9 · walk-forward fold 11/12 · last trained 06:10

VisionGuard

In useBrand protection · SaaS

Brand asset protection for agencies and the talent they represent. VisionGuard crawls 44 platforms across social, marketplaces, stock, design, NFT and ad networks, detects unauthorized use of a brand's imagery, automates DMCA takedowns and produces court-ready forensic evidence with RFC 3161 timestamps and a hash-linked chain of custody.

Detection pHash triage → OpenCLIP embeddings with ANN search in pgvector → XGBoost classifier  ·  Platform FastAPI + Celery, Supabase/pgvector, Redis, Cloudflare R2, Clerk, React on Vercel · schema-per-tenant Postgres with agency white-label

VisionGuard · Match reviewliveapp.visionguard.io/matches?status=pending
Protected assets1,284
Crawl targets44all healthy
Pending review37
Takedowns this week19
MatchPlatformConfidenceStageEvidenceStatus
campaign_ss26_look04.jpgInstagram0.97CLIP + XGBsealed · RFC 3161DMCA sent
editorial_mg_0112.pngEtsy listing0.94CLIP + XGBsealed · RFC 3161Removed
talent_ak_headshot_v2.jpgTikTok ad0.89CLIP + XGBpending sealReview
lookbook_p07_crop.jpgShutterstock0.61pHash only—Low signal

GrantGuard

Built end-to-endCompliance · RAG

Federal grants compliance platform. Each expense is run through a three-step allowability evaluation using retrieval over the specific award's documents, scoped by award ID, so the answer cites the clause that governs it. Every decision lands in an append-only, HMAC-signed audit log.

Stack FastAPI · Next.js 14 with shadcn/ui · schema-per-tenant Postgres · AES-256-GCM encryption at rest

GrantGuard · Expense reviewgrantguard.app/awards/NIH-R01-4471/expenses/E-2093
E-2093 · Conference travel, Boston · $1,842.50Allowable
1
Cost principle check2 CFR 200.475 travel costs: reasonable, consistent with institutional policy. Lodging within GSA per diem.
2
Award-specific termsNotice of Award §IV.2 permits domestic travel for dissemination. No prior approval required under $5,000.
3
Budget period & allocationWithin budget period 03 · charged to travel line (remaining $6,210 after this expense).

Governing clauseretrieved · award_id scoped

"Travel costs are the expenses for transportation, lodging, subsistence, and related items incurred by employees who are in travel status on official business of the non-Federal entity…"
NoA_NIH-R01-4471.pdf · p.4 · chunk 0117 · cosine 0.91

Audit logappend-only · HMAC

#4,118 eval.complete E-2093 allowable sig=9f3a…c41e#4,117 retrieve.award NIH-R01-4471 6 chunks sig=22b0…8d71#4,116 eval.start E-2093 user=jmoreno sig=d7e5…10af

Clawbot

LivePrediction markets · Trading

Automated trading system for binary prediction markets, extended to US equity contracts on Polymarket. Kelly-criterion position sizing, a typed Pydantic model layer for every market object, and a kill-switch architecture that halts the book on any anomaly. Current work: cross-venue arbitrage between Kalshi and Polymarket, with an adapter and a market-matching layer that maps Kalshi tickers to Polymarket condition IDs.

Stack Python · Pydantic · venue adapters · fractional Kelly sizing · hard kill switch

clawbot · operator consoleprodtmux: clawbot-prod · 0:engine 1:arb 2:risk
Kill switchARMED
Open positions7
Book exposure41.6%of cap
Kelly fraction0.25
12:41:07 engine tick venues=[polymarket,kalshi] latency p50=142ms p99=388ms12:41:07 arb match KXFED-26OCT-T425 ↔ 0x8c1f…a9e2 spread=+3.4c edge=2.1% size=$480 (f*=0.25)12:41:08 risk ok exposure 41.6% < 60% cap · max_single 9.8% < 15%12:41:08 engine fill BUY YES 0x8c1f…a9e2 qty=1,250 @0.611 SELL YES KXFED-26OCT-T425 qty=1,250 @0.64512:41:12 arb match KXCPI-26NOV-T31 ↔ 0x2d77…40b1 spread=+1.1c edge=0.6% skip (below 1.5% floor)12:41:30 risk watch price_jump 0x5a09…ee3c Δ=-14.0c in 9s → quarantine market, no new orders12:41:30 engine note kill_switch conditions: 0/4 triggered (pnl_dd, venue_desync, model_stale, api_err)

MarketMind

ResearchUS equities · Reinforcement learning

Reinforcement-learning trading system for US equities. Sequence models (LSTM and Transformer) feed state to SAC and PPO agents, with Avellaneda–Stoikov quoting for the market-making regime. Evaluated strictly with walk-forward backtesting; no in-sample results reported.

Stack stable-baselines3 · PyTorch · Polygon.io market data · Alpaca execution

marketmind · walk-forward run 2026-10-11runs/wf_sac_lstm_v3/

Walk-forward foldstrain 18m → test 3m, rolling

F1F2F3F4F5
trainout-of-sample testrunningqueued
$ mm run --config wf_sac_lstm_v3.yamlfold 1/5 train 2023-01→2024-06 test 2024-07→09 donefold 2/5 train 2023-07→2024-12 test 2025-01→03 donefold 3/5 train 2024-01→2025-06 test 2025-07→09 donefold 4/5 train 2024-07→2025-12 test 2026-01→03 training ep 1,840/4,000 policy=SAC obs=LSTM(64) quoting=avellaneda-stoikov γ=0.1 reward: inventory-penalised pnl · slippage model: polygon L2fold 5/5 queued
Client work

Talent agency, New York

Sector
Talent & creator management
Problem
Client imagery and headshots reused without license across social and marketplaces; takedowns handled by hand
Engagement
VisionGuard pilot, started Sept 2026
Status
Live, monitoring the agency's roster

The agency's team was finding stolen images by searching manually and filing DMCA notices one by one, with no record that would hold up if a dispute escalated. We onboarded their roster into VisionGuard, ran detection across all 44 crawl targets, and routed confirmed matches through the automated takedown and evidence-sealing flow.

Early-stage matches are reviewed with the agency rather than fully automated, so the classifier is trained on their actual false positives before it runs unattended. That's the model for every Operon engagement: ship, watch it with the client, then let it run.

Products I'm shipping

I also build and launch my own consumer products end to end, from the ML core to the App Store listing. It's how I stay honest about what "production" means, and it's why Operon can ship yours.

Borealis

TestFlightiOS · Wellness

A flicker-light and sound stimulation app for iPhone: 40 Hz gamma and 10–15 Hz alpha modes for cognition, mood and sleep. Built with a cofounder; usage data is collected for research under a secure academic research cloud.

Stack Expo / React Native · research data pipeline

Trueline

BuildingiOS · Meditation

Meditation, affirmations and manifestation in one app, with AI-personalized audio sessions. Distribution runs through partnerships with creators who become the faces of the product.

Stack Expo / React Native · AI audio generation · creator partnership model

Savant

Live storeStreetwear · Shopify

A clothing brand built from the ground up: custom Dawn-based Liquid theme, print-on-demand fulfillment through Tapstitch, AI on-model product photography and short-form video ads. First drop: the Roots collection.

Stack Shopify Advanced · custom Liquid theme · Tapstitch POD

Background
  • Now
    Founder, Operon Labs

    AI automation studio for finance and revenue operations, New York.

  • Previously
    Quantitative Analyst, BNY Mellon

    Derivatives pricing.

  • Previously
    Silver Slate Group

    Built in-house automation systems of various kinds across the firm's operations.

  • Education
    Baruch College, Zicklin School of Business

    Dual BBA: Statistics & Quantitative Modeling; Operations Management & Political Science.

Contact

Working on finance or revenue operations and tired of the manual layer?

Tell me what breaks every month-end. I'll tell you whether it's a two-week build or something you shouldn't automate at all.

Book a call
Emailstevennakhwal@gmail.com