Software engineer building AI systems and the infrastructure underneath them. Software Engineer at UF Health · M.S. Computer Science, University of Florida · Y Combinator Startup School 2026
Website: rohanreddy79.github.io · LinkedIn · jakkamrohan79@gmail.com
I work at the layer where machine learning meets systems engineering: training loops, retrieval pipelines, agent runtimes, inference servers, and the consensus, indexing and gateway code that production platforms depend on. In industry I have shipped an IRB-regulated healthcare data platform used by 100+ NIH-funded researchers, a zero-to-one serverless video-analysis engine, and ad-platform features for enterprise customers. The repositories below are written from scratch, tested in CI, benchmarked with reproducible commands, and documented with honest roadmaps.
| Project | What it is | Highlights |
|---|---|---|
| tinygpt | A GPT trained from scratch in PyTorch: byte-level BPE tokenizer, decoder-only transformer, KV-cache generation | 870K-parameter model trained on a laptop CPU in 147 s to 2.86 val loss; KV cache decodes 2.8× faster than recompute; 64 tests |
| ragkit | Hybrid retrieval-augmented generation: BM25 + dense embeddings fused with reciprocal rank fusion, reranking, citations, CLI and HTTP API | recall@3 0.98 and MRR 0.94 on a 45-question eval corpus; BM25, dense and hybrid compared against a measured noise floor; 55 tests |
| agentloop | A minimal agent runtime: typed tools with generated JSON schemas, budgets, parallel tool calls, structured outputs, JSONL tracing, Claude provider | Anthropic provider tested against a mocked SDK client; 8 parallel tools finish in 51 ms vs 413 ms serial; 153 offline tests |
| evalkit | An evaluation harness for LLM apps: declarative tasks, deterministic metrics and LLM-as-judge, response cache, bootstrap confidence intervals, paired run comparison | paired bootstrap and sign test detect a +6.5-point change (95% CI +1.0 to +12.0) that overlapping intervals would hide; 63 tests |
| batchserve | An inference server with dynamic batching, bounded queues with backpressure, Prometheus metrics and a Docker image | 17× throughput from batching (94 → 1,641 RPS, p50 648 → 36 ms); classifier at 96.7% on held-out templates; Docker image; 51 tests |
| nanograd | A tensor autograd engine and neural-network library on NumPy, every gradient verified numerically; trains an MLP and a CNN | 30 differentiable ops, 75 finite-difference gradient checks; CNN reaches 98.3% on 8×8 digits in 1.7 s of CPU training; 156 tests |
| Project | What it is | Highlights |
|---|---|---|
| raftkv | Raft consensus and a replicated key-value store in Go, with linearizable reads and a deterministic simulated network for partition and crash tests | Figure 2 Raft with ReadIndex reads and exactly-once client sessions; 20 seeded partition schedules under the race detector; ~300 ms failover |
| hnswgo | An HNSW approximate nearest-neighbour index from scratch in Go: persistence, deletes, concurrent search, HTTP service, recall/QPS benchmark | recall@10 of 0.999 at 35k queries/s single-threaded (12× brute force) on 20k×64 vectors; one allocation per query; CRC-checked persistence |
| llm-gateway | An OpenAI-compatible LLM gateway in Go: virtual keys, rate limits, retries, provider fallback, circuit breakers, response cache, streaming translation, cost accounting | 65k requests/s at p99 1.7 ms through the full middleware chain; a 10%-failing upstream served 3,000 requests with zero client-visible errors; 87 race-clean tests |
| mcp-toolbox | A sandboxed Model Context Protocol server in TypeScript: read-only SQLite, allow-listed filesystem and HTTP tools, with tests that prove each boundary | 8 tools over stdio and Streamable HTTP; 18 sandbox checks each mapped to a test; zero native dependencies via node:sqlite; 100 tests |
Coursework and smaller projects from 2022–2025, each brought up to the same standard in October 2026: a real README, MIT license, pinned dependencies, the known bugs fixed, and a CI workflow that builds or tests it on every push.
| Project | What it is | Language |
|---|---|---|
| distributed-computing-model | Gossip and push-sum convergence simulated over line, full, 2D and imperfect-3D actor topologies with Akka.NET | F# |
| chord-dht | Chord peer-to-peer lookup: 160-bit SHA-1 ring, finger tables, stabilisation; measures about ½·log₂N hops per lookup after the routing fix | F# |
| concurrent-client-server | Concurrent TCP calculator server and client with a small error-code protocol | F# |
| GatorLibrary | Library simulator on a hand-written red-black tree with min-heap reservation queues; golden-output test | Python |
| Simple-File-Transfer-Chat | Peer-to-peer terminal chat with inline file transfer over raw TCP sockets | Python |
| password-cracking-attacks | Security coursework: brute-force, dictionary, rainbow-table and divide-and-conquer attacks on toy hashes | Python |
| log-report-fixed | A Terminal-Bench 2 task with its oracle and verifier, checked end to end in Docker | Python |
| Django-deploy | Django + DRF polling app with a JSON vote API, fixed and covered by tests | Python |
| streamlit-nyc-collisions | Streamlit dashboard over NYC motor-vehicle collision data: maps, hourly hexagons, top streets | Python |
| Data-Science-and-analytics-Projects | Seven notebooks: regression, k-means, retail and IPL EDA, decision trees, ARIMA on SENSEX, COVID maps | Jupyter |
| Wetland-Detection-ML | Wetland vs dry-land classification from spectral bands: trees, logistic regression, SVM, bagging, random forest | Jupyter |
| sklearn-clustering-lab | K-means, agglomerative and DBSCAN compared on synthetic blobs (re-executed on current scikit-learn) | Jupyter |
| Random-Clusters | The same clustering lab with the written analysis | Jupyter |
| TASK-Check | Task dashboard and CLI with a Node test suite | JavaScript |
| assessment | Package-sorting exercise with 14 unit tests | Python |
| spidr-airfryer-form | A small React form built for a take-home exercise | JavaScript |
| Portfolio | Previous portfolio site (React, Vite, Tailwind); superseded by rohanreddy79.github.io | JavaScript |
- Every repository has a test suite and a CI workflow; nothing is merged red.
- Every number in a README comes from a run on a stated machine, with the command to reproduce it.
- Hosted models are optional: tests and demos run offline with deterministic fakes.
- Each README explains the design decisions and the trade-offs, not only the usage.
- UF Health · Software Engineer · Jan 2026–present · Scoped and shipped an IRB-regulated full-stack data platform (TypeScript, React, Python, AWS) for 100+ NIH-funded researchers: 99.9% uptime, interactive visualisation over 10M+ rows of healthcare data, 40% less manual retrieval and analysis time, demoed to NIAAA program leadership.
- FilmTrends · Founding Engineer · Oct 2025–Jan 2026 · Built the product from zero: a serverless video-processing engine on AWS Lambda analysing 10,000+ social accounts in parallel (latency 5 min → 30 s), an ML viral-prediction model over 50+ video features at 78% accuracy, infrastructure as code with Terraform and CloudFormation.
- Koddi · Software Engineer · Jun–Oct 2025 · Client-facing ad-platform features (React, TypeScript, GraphQL, Go) for Booking.com, Kroger and Grubhub; Go REST APIs and real-time pacing dashboards that cut customer-visible data discrepancies 45%.
- University of Florida · Software Engineer · Jan 2023–May 2025 · High-availability admin dashboard (Java, Spring Boot, React) consolidating 8 workflows at 99.9% uptime; Jenkins and Docker CI/CD that tripled release frequency.
- Vera · Voice AI agent platform for regulated industries. Agents whose actions are verifiable by design: a compliance finite-state machine constrains behaviour in real time, and a cryptographic audit ledger makes every action tamper-evident.
- Layman · AI news summariser live on the App Store. Cross-platform MVP shipped in 48 hours; production RAG backend (pgvector, Prisma) on an AWS Bedrock/SageMaker pipeline turning jargon into plain-language summaries.
- AppleTune · macOS menu-bar app in Swift for per-app volume, multi-device output, audio routing and a 10-band EQ.
- Languages: Python, TypeScript, Go, JavaScript, Java, SQL, Swift, C++
- AI / LLM: RAG pipelines, agents and tool use, evaluation frameworks, guardrails, vector search (pgvector), OpenAI and Anthropic APIs, MCP, AWS Bedrock and SageMaker, LangChain
- Full-stack: React, Next.js, React Native, Node.js, FastAPI, Flask, GraphQL, REST, OAuth2/JWT
- Data and infrastructure: PostgreSQL, Redis, Kafka, Spark, Airflow, Snowflake, ETL pipelines
- Cloud and DevOps: AWS, GCP, Docker, Kubernetes, Terraform/CloudFormation, Jenkins, CI/CD, production monitoring
- University of Florida · M.S. Computer Science · Dec 2024
- Manipal University · B.Tech Computer Science · May 2023
- Y Combinator Startup School · 2026 cohort