SUKHMAN VIRK
Full-Stack Software Engineer · Python · TypeScript/React · Production AI & Agent Systems
EDUCATION
EXPERIENCE
- Forward-deployed engineer on enterprise engagements, scoping requirements with client stakeholders and authoring the solution design documents they sign off on.
- Leading development of a client earth-observation platform on Sentinel-2 and Landsat imagery, directing 4 engineers toward self-serve onboarding of client data sources for analysis.
- Designed and built an AWS-funded GenAI pipeline that localizes English micro-dramas into Hindi and Egyptian Arabic, processing 61 episodes end to end in the proof of concept.
- Benchmarked every model head-to-head on hand-labeled data and held LLM speaker naming to pre-set accuracy and cost bars, with candidates running at under $0.01 per episode.
- Sole engineer on a full-stack custody platform in Django and React, shipped in under 4 weeks, tracking ~27,000 cataloged space objects and auto-assigning each object's geographic command from its orbit.
- Built the custody platform's AI layer, where Claude answers questions through 5 MCP tools that enforce role-based access, over Bedrock hybrid vector and keyword search.
- Owned natural-language search for a multi-tenant weather and satellite data marketplace, turning a sentence into PostGIS queries scoped to each customer's subscription plan.
- Built the UI and cost-review engine for Minfy's internal Claude Code platform, attributing AI spend per engineer, project, and delivery phase without storing any prompt content.
- Built the React admin console and Go backend for a white-label, multi-tenant smart-lock platform on AWS Lambda, DynamoDB, and Cognito, keeping each brand's users and devices isolated.
- Built the smart-lock platform's dashboard analytics, serving each metric from one indexed DynamoDB query instead of a log scan.
- Built a C++ WebRTC peer-to-peer app for a Fortune 500 company, transferring images, video, and messages at up to 4K.
- Automated ICE negotiation through Google STUN so peers connect across networks without manual setup.
- Built a full-stack AES-256 encryption service in Next.js and Java protecting 50,000+ user records.
- Automated decryption-key retrieval through AWS KMS and integrated Okta 2FA.
PROJECTS
A two-sided market maker on Kalshi BTC hourly binaries, quoting the full strike ladder in under 50ms. It ran live with real capital across 10k fills ($10K+ traded), and its losses traced to adverse selection during volatile moves.
- Built a participation gate that pulls quotes during information events, after tracing the losses to adverse selection; it halted quoting 500+ times in shadow trading.
- Modeled queue fill probability with a survival model trained on 1M+ order events, handling orders cancelled or amended before they filled.
- Priced adverse selection per fill with a gradient-boosted markout model, so quotes target edge net of expected markout rather than raw spread.
- Set spread, size, and hold time with an inventory-aware profit optimizer.
- Centers quotes on a reservation price in log-odds space rather than the Avellaneda-Stoikov closed forms, whose spread and skew terms degrade on binary contracts near the tails.
- Built the execution stack in Python and Redis: signed exchange auth, amend-first order management, and fill reconciliation against exchange state to prevent phantom inventory.
An options research platform in Next.js, Flask, and PostgreSQL with a daily ETL of equities, options chains, and Treasury yields.
- Black-Scholes IV inversion (Newton-Raphson), Greeks, and spline-smoothed implied volatility surfaces.
- LRU caching on the iterative numerical solvers for 4x faster queries.
- Co-published cuda-kernels on PyPI: shared-memory CUDA kernels for autocorrelation and sum reduction with a NumPy fallback.
App repo · PyPI · Kernels repo
A Dixon-Coles team-strength model, a multi-period MILP squad optimiser, and a backtest harness that gates every change on held-out seasons. The MILP cleared its pre-registered bar on a held-out season (+56.3 points over greedy, 5 of 6 seeds ahead). Benchmarks weekly against managers with repeated top-1% finishes, read live from the FPL API.
End-to-end NLP pipeline on 14,600 tweets: TF-IDF features with SVM and Logistic Regression for 3-class sentiment, K-means with PCA clustering validated by Silhouette scoring, and LDA topic modeling to surface latent complaint themes.
Real-time Android app in Java rendering friends' live locations from persistent location streams over a SQL-backed friend graph, with a full Robolectric, JUnit, and Espresso test suite on GitHub Actions CI/CD.
SKILLS
Languages: Python, TypeScript, SQL, Go, JavaScript, C++, Java, C
Frameworks & Libraries: FastAPI, Flask, Django, React, Next.js, Node.js, PyTorch, TensorFlow, Pandas, NumPy, CUDA, Redis
Cloud & Data: PostgreSQL/PostGIS, AWS (Bedrock, Lambda, ECS, App Runner, S3, DynamoDB, RDS, Cognito, KMS, OpenSearch, CloudFront, CloudWatch), Terraform, Docker, GitHub Actions
Domains: LLM Systems & Evaluation, RAG, MCP/Agentic Systems, Machine Learning, NLP, Market Microstructure
AI-Assisted Development: Claude Code, Cursor, GitHub Copilot
LEADERSHIP
- Built a QR check-in system that cut queue wait times by 50% and eliminated unauthorized entry across 10+ events.
- Maintained a website serving 7,000+ monthly visitors.