Ankit Kumar

FunctionGemma Phone Assistant — on-device function calling

• Built a 4,000-example English/Hinglish function-calling benchmark (20 phone tools; single, parallel and no-call requests) with leakage-safe splits; published dataset and model on Hugging Face (481 model downloads).
• Full-parameter fine-tuned FunctionGemma-270M with DDP on 2× T4 GPUs: held-out exact match 22.3% → 75.4%, tool selection 31.6% → 93.0%.
• Designed symbolic scoring (exact match, tool selection, no-call precision/recall) and reported the trade-off alongside the gains: no-call recall fell 90.4% → 79.8%.

Gemma 4 Calculus SFT — how much data does fine-tuning need?

• Ran a controlled QLoRA learning-curve study on Gemma 4 E2B with nested training sets (128–4,096 examples), validation-only model selection, a single-use test set and SymPy grading.
• Selected adapter lifted reserved-test accuracy 32.1% → 50.4% (McNemar p ≈ 1e‑5); also surfaced regressions from small runs and negative transfer on an unseen composition.

Kyoto — AI browser agent for developers

• Built a Chrome side-panel agent for any OpenAI-compatible provider (OpenRouter, DeepSeek, Ollama, ...) that drives the browser through ~100 tools: DOM, network capture, cookies, debugger, shell.
• Implemented a streaming tool-calling loop (up to 12 round-trips per turn) with reasoning passthrough, MCP servers over JSON-RPC, per-tab sessions, cross-session memory and per-message cost tracking.

Hide-and-Seek Replication — emergent tool use for $3

• Replicated OpenAI's multi-agent hide-and-seek (ICLR 2020) through the chase and fort-building phases for $3.10 of GPU time: PufferLib PPO at 1.2M agent-steps/s, 20B steps.
• Wrote the C environment (~111K steps/s/core; occluded vision, lidar, grab/lock); scripted-bot certification caught six env-breaking bugs before any GPU spend.

6-Link Inverted Pendulum — RL control from first principles

• Trained PPO to balance a 6-link pendulum on a cart: 7-DOF Lagrangian dynamics in C, 1B steps in ~6 min on one RTX 4090 (3.4M steps/s, ~$0.50); 100/100 eval episodes at full horizon.
• Certified controllability with an LQR study before training; then solved free-joint swing-up (100/100) with a self-paced reverse curriculum for ~$2 of GPU.

Associate Software Development Engineer

Client: Goldman Sachs (US) — Podium, private wealth platform Aug 2024 – Present

• Led system design (HLD/LLD) for Podium and drove architecture decisions in bar-raiser forums with zero regressions.
• Owned end-to-end delivery of Spring Boot microservices serving private wealth users at 99.9%+ availability.
• Built property-driven, configurable GraphQL/REST APIs that multiple divisions can adopt without code changes.
• Added Redis and in-memory caching with startup hydration, cutting average latency ~40% and stabilizing p99.
• Integrated multiple downstream systems with Resilience4j (retries, circuit breakers), improving fault tolerance.
• Established production observability with Prometheus and Grafana, reducing MTTR ~30%.
• Ran load and performance testing with Gatling and enforced unit/integration testing standards.
• Mentored engineers and set backend development standards across the team.

Client: Lloyds Bank — banking analytics platform Nov 2023 – Jul 2024

• Designed and built a Spring Batch ETL pipeline that extracts data from Cloud SQL, converts it to Parquet and archives it to Google Cloud Storage as a backup.
• Integrated BigQuery for efficient transaction analytics and reporting.

Research Intern (Remote)

• Collaborated on behavioral studies of content consumption on platforms such as YouTube and Facebook.
• Built automated data-collection and analysis tools (FastAPI, React, SQLite), improving research efficiency.
• Implemented a backup solution for c2c data transfer that ensured data integrity.

SOS — Search Over Screenshots

• Users upload images to their account; Tesseract OCR extracts the text and images are categorized by content.
• Implemented full-text search over extracted text with Elasticsearch.

Code-Collab

• Built a real-time code editor where multiple users work on the same code at once, with multi-language support and themes.
• Real-time broadcast messaging between users over WebSockets.

Maulana Azad National Institute of Technology (MANIT), Bhopal

• 1st prize, IEEE Ideathon
• 1st prize, NITS Hack 5.0