ShreyansChowdry

A relational database engine with Git semantics: commit, branch, merge and time-travel your data.

unit tests

126

branches, 2,000+ versioned rows

5

pages, LRU buffer pool

4 KB

• Built a Git-style relational database engine from scratch with immutable commits, copy-on-write branching, time-travel queries and conflict-aware three-way merge over a commit DAG.
• Wrote the storage layer (LRU buffer pool, B+ Tree indexes, WAL with ARIES-style redo recovery, mark-and-sweep GC), backed by 126 unit tests and benchmarked against SQLite.

• Python
• FastAPI
• Next.js
• TypeScript

Code on GitHub

A compiler for a C subset whose register allocator learns which paths are hot from a profiling run, then spends registers where they save the most cycles.

cycles vs linear scan (geomean)

-20.4%

spill memory traffic vs linear scan

-80%

benchmark runs verified against the interpreter

50/50

• Built, in a team of 5, a full compiler for a C subset: DFA lexer, LL(1) recursive-descent parser, semantic analysis, pruned SSA, an SCCP/GVN/DCE optimiser and a cycle-counting RISC simulator; I owned the lexer, parser, AST and IR generation.
• Its profile-guided register allocator (per-value MDPs solved by policy iteration, sealed-bid auctions on the chordal interference graph) cut cycles by 20.4% and spill memory traffic by 80% vs linear scan across 50 benchmark runs; written up as an IEEE-format paper.

• Python
• JavaScript

Code on GitHub

A job queue built on Redis primitives that survives crashed workers and flaky jobs.

jobs completed with 30% injected failures

99.2%

ingest throughput

410/s

jobs lost to dead workers

0

• Designed a distributed job queue on Redis sorted sets, hashes and Pub/Sub with priority scheduling, delayed execution and PostgreSQL as the durable source of truth.
• Added heartbeats, a reaper that re-queues jobs from crashed workers, backoff retries and a dead-letter queue; in a 500-job load test with 30% injected failures, 99.2% completed at 410 jobs/s ingest.

• Python
• Redis
• PostgreSQL
• Docker

Code on GitHub

Trip planning solved as a routing problem: which places, which day, what order, inside opening hours.

travel time per day

-45%

opening-hour conflicts (v1: 37% of days)

0

p95 solve time

<25 ms

• Modelled trip planning as a Team Orienteering Problem with Time Windows and built an Iterated Local Search solver (greedy insertion, 2-opt/Or-opt, inter-day relocate and swap).
• Across 405 benchmark trips, cut daily travel time by 45% and eliminated the opening-hour conflicts that broke 37% of the old planner's days, at p95 solve time under 25 ms.

• JavaScript
• React
• Vitest
• Vercel

Code on GitHub

Ask a repository questions in plain English and get answers grounded in the code, with citations.

top-15 retrieval hit rate

93.3%

average retrieval latency

312 ms

vector + keyword fusion

RRF

• Built hybrid retrieval fusing pgvector cosine search and Postgres full-text search via Reciprocal Rank Fusion, with 1-hop import-graph expansion over tree-sitter code chunks.
• Reached a 93.3% top-15 retrieval hit rate on hand-annotated queries at 312 ms average latency; streamed answers over SSE with inline file citations.

• FastAPI
• pgvector
• Claude API
• Next.js

Code on GitHub

Six cooperating agents that search the literature, read the PDFs and write up the research gaps.

agent nodes in a LangGraph state machine

6

LLM tokens via reference stripping

-40%

live paper APIs (OpenAlex, Semantic Scholar)

2

• Orchestrated a 6-node LangGraph agent pipeline that searches OpenAlex and Semantic Scholar, parses full-text PDFs, extracts structured findings and synthesizes citation-backed research gaps.
• Stripped bibliographies during PDF parsing to cut LLM token usage by up to 40%; shipped with Docker Compose and GitHub Actions CI.

• LangGraph
• Gemini
• FastAPI
• PostgreSQL

Code on GitHub

Predicting which passing dies on a wafer will fail final test, and explaining why.

AUC-PR over baseline

+7.8%

recall on latent failures

+8.8%

positive class: heavily imbalanced

4.2%

• Predicted latent die failures on a highly imbalanced (4.2% positive) wafer dataset by gradient boosting over 549 parametric, KD-tree spatial and PCA-compressed features with wafer-grouped CV.
• Lifted AUC-PR by 7.8% and recall by 8.8% over the baseline model, confirmed by a 2,000-resample wafer-level bootstrap; explained predictions with per-die occlusion attribution.

• scikit-learn
• SciPy
• pandas

Code on GitHub

Turns noisy kernel and system logs into correlated incidents with a likely root cause.

log sources correlated

3+

simulated failure scenarios

10+

• Python
• FastAPI
• Next.js

Code on GitHub