- Built a GitHub App that diagnoses failed CI runs on pull requests with an LLM (Claude or Gemini, typed structured output) and posts the root cause plus a fix, shown only if it applies to the PR's head commit; locally it also runs the tests on the patched code and hides fixes that still fail.
- Designed a broker-less Postgres job queue (SKIP LOCKED) with webhook signature checks, deduplication, retries with backoff and crashed-job recovery; a concurrency test shows exactly-once claiming across 8 workers.
- Wrote a log extractor covering ~30 failure patterns across 8 toolchains that sends ~0.3% of each log to the model and kept the key error line in 100% of held-out cases; on a 30-case eval, Gemini 3.5 Flash got the failure type right in 25 of 26 answered cases and named the file to fix in 17 of 19.

Open to new-grad SWE roles · MS graduating Dec 2026
Shibin Biji
Thomas
Software Engineer. Hands-on experience in AI/ML, full-stack and cloud backend engineering, from shipping production systems at LinkedIn and Accenture.
prev AI/ML & SWE Intern @ LinkedIn · SWE @ Accenture · MS @ SJSU ’26
- Feed ranking & LLM pipelines
- AI/ML
- Feed ranking & LLM pipelines · LinkedIn
- full-stack software engineer
- 3.5 yrs
- full-stack software engineer · Accenture
- open-source AI engine I contribute to
- 17.8k★
- open-source AI engine I contribute to · rocketride

/ about
A bit about me
I'm a software engineer finishing an MS at San Jose State (December 2026). At LinkedIn I worked on the feed recommendation system, shipping ranking-model improvements in PyTorch and TensorFlow (11 of 12 AUC wins), along with RAG and LLM evaluation pipelines.
Before that I spent three and a half years at Accenture as a full-stack engineer, building React and TypeScript apps on Go, Java and Python microservices on AWS, and raising platform availability from 98.2% to 99.7%.
I also contribute to rocketride-server, a 17.8k-star open-source C++/Python AI pipeline engine, fixing bugs in its CLI, LLM provider integrations, and database and graph nodes.
what I work on
- ML ranking & recommendation systems (PyTorch, TensorFlow)
- LLM apps, RAG & evaluation
- Full-stack web apps (React, Next.js, TypeScript)
- Backend & distributed systems (Go, Python, Java)
- Cloud infrastructure (AWS, GCP)
/ experience
Where I've worked
Software Engineer Intern · LinkedIn
May 2026 – August 2026
Sunnyvale, CA, USA
- Engineered automated cloud telemetry, workflow automation, and audit monitoring pipelines on AWS (SNS → Firehose → S3, Lambda) and GCP using Go and Python, reducing manual reporting effort by 90%.
- Delivered production ranking-model optimizations for the feed recommendation system in PyTorch and TensorFlow across concurrent pipelines, achieving 11/12 AUC wins (up to +0.20% Click) and 11/12 Normalized Entropy (NE) improvements.
- Developed high-throughput serverless microservices and backend APIs in Go and Python, optimizing stream processing and event feedback loops to cut latency from 120ms to 85ms.

Software Engineer · Accenture
July 2021 – December 2024
Pune, India
- Architected scalable, serverless microservices and backend transaction systems in Go, Java, and Python, translating client requirements into production systems that lifted platform availability from 98.2% to 99.7%.
- Developed client application web interfaces and integrations using React.js, TypeScript, and RESTful APIs, designing gRPC contracts and GraphQL schemas with Protobuf to standardize data flow.
- Optimized relational database schemas and complex queries across PostgreSQL and SQL, tuning indexing and connection pooling to cut backend API response times from 450ms to 280ms.
/ open source
Contributing upstream
An open-source, multithreaded C++17/Python engine that executes AI workflows across 130+ Python nodes, 15+ LLM providers, and databases.
- PR #2679Merged
Fix bare --modelserver port being read as a hostname
Found and fixed a bug where a bare port in --modelserver (e.g. --modelserver=5590) was read as a hostname, so the Python client tried to connect to ws://5590:5565/models and failed.
PythonWebSocketsCLICross-platform CI - PR #2680In Review
Send and cap Gemini output token limits
Fixed the Gemini LLM node so the configured output-token limit is actually sent to Google and saved where it is read back.
PythonGemini APILLM ProvidersUnit Testing - PR #3276In Review
Point agents from a disabled graph execute tool to the read tools
Fixed issue #2157: with execute disabled (the default), an agent asking a graph database a read-only question got only "execute tool is disabled", which never said what to use instead. The error now points to get_data (and query on FalkorDB), keeping the original wording so existing checks still match.
PythonGraph DatabasesFalkorDBNeo4jAI AgentsUnit Testing - PR #3277In Review
Name the missing variable when a database or graph credential is unfilled
Part of issue #2712: a database or graph node whose password or token was still an unfilled ${ROCKETRIDE_*} placeholder sent that literal text to the server and failed with an authentication error that never named the variable. The run now stops at start and names the missing variable, and saving the node shows a warning instead of contacting the server.
PythonPostgreSQLMySQLClickHouseNeo4jRedisUnit Testing
/ projects
Things I've built
- Built a multi-service platform in Go, Java and Python: a Go gateway ingests client events over gRPC and REST with per-tenant JSON Schema validation (~78K events/s from one client) into an S3 data lake, and a Java GraphQL API versions schemas and rejects breaking changes.
- Orchestrated a nightly Airflow pipeline where Spark compacts raw JSON into partitioned, de-duplicated Parquet and Trino builds rollups with row-count reconciliation, cutting dashboard query latency by 70% over 2.35M events.
- Built a RAG assistant with hybrid pgvector + full-text retrieval, cited streaming answers and feedback-aware ranking (0.89 hit@3 on a 26-question eval), plus a React + TypeScript dashboard for real-time telemetry.
- Built a crash-safe key-value storage engine as a log-structured merge tree, the design behind LevelDB, RocksDB and Cassandra: checksummed write-ahead log, SSTables with Bloom filters and block index, and levelled background compaction.
- Added MVCC snapshots and range scans that stay consistent while writers and compaction keep running; crash tests kill a real process at every step of every disk operation.
- Benchmarked against RocksDB and SQLite on 1M keys: level with RocksDB on random writes (381K vs 386K ops/s) and 1.6x faster on in-memory point reads.
- Built a tool that finds regular expressions vulnerable to catastrophic backtracking (ReDoS) in JavaScript, TypeScript, Python and Ruby code, using a step-counted model of how each engine backtracks.
- Proves every finding by generating the attack input and running it on the real Node, Python and Ruby engines at growing lengths; 100% recall and 0 false positives on a self-verifying corpus.
- Ships as a CLI, pre-commit hook and GitHub Action with SARIF output, so it can gate CI.
- Turns hospital discharge papers into a plan patients can follow: an LLM extracts medicines, appointments and warning signs with source quotes, and code verifies every item, drug name and number against the original paper.
- Generates plain-language plans in 8 languages with code-checked safety rules for duplicate ingredients, stopped medicines and daily dose limits.
- Runs daily check-ins over SMS (Twilio) and Telegram, escalating worrying replies to a care-team dashboard; covered by 804 automated tests, including a 336-case safety evaluation.
🔊 Narrated demo, 1:45 · unmute for the voiceover · Watch on YouTube ↗
- Upload a Zoom, Meet or Teams transcript and get a summary, decisions and action items, each linked to the exact transcript timestamp; the model cites segment numbers so timestamps cannot drift.
- Ask questions across all meetings and get streamed answers (server-sent events) with numbered citations, retrieved from an SQLite FTS5 BM25 index.
- Retrieval scores hit@3 = 1.00 on a 20-question eval that runs in CI; fully Dockerized with GitHub Actions.
- Team project built at Cal Hacks 13.0: a desktop tutor for Windows and macOS that reads your screen, points at what matters, draws hand-drawn marks and working beside the problem, and talks you through it step by step.
- Fastify orchestrator streams Claude's answer as structured steps, so voice narration starts about 3 seconds after a question while the rest of the lesson is still being generated.
- Cartesia text-to-speech drives the timing: each step's pointer and drawings appear when its audio starts, keeping voice and visuals in sync.
- Guided lessons for math, chess, software and languages watch the screen for the user's next action and check it before continuing, backed by a 13-case accuracy eval set and 380+ unit tests.
More projects
Autonomous AI Agent & Serverless Workflow Engine
Engineered an autonomous AI agent framework and serverless workflow engine in Go, Python, and Rust, utilizing custom RAG pipelines, LLMs, and Kafka to automate complex manual handoffs.
/ skills
Tools I reach for
Languages
09Frameworks & APIs
09Databases & Cloud
07AI & LLMs
10DevOps & Testing
08Data & Distributed Systems
08/ education
Education
San Jose State University
January 2025 – December 2026 (Expected)
Master of Science in Applied Data Intelligence
San Jose, CA · 3.5 GPA
/ contact
Let's talk
I'm graduating in December 2026 and looking for new-grad software engineering roles. If you're hiring, or just want to talk about something I've built, my inbox is open.