Skip to content
View G26karthik's full-sized avatar
54 followers · 6 following

Achievements

Achievement: StarstruckAchievement: Pull Sharkx2Achievement: YOLOAchievement: Pair ExtraordinaireAchievement: Quickdraw

Achievements

Achievement: StarstruckAchievement: Pull Sharkx2Achievement: YOLOAchievement: Pair ExtraordinaireAchievement: Quickdraw

Highlights

  • Pro

Organizations

@openMF @fossasia

Block or report G26karthik

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
G26karthik/README.md
G. Karthik Koundinya, Applied AI Engineer and Researcher. I build and research LLM systems, quantization, quantum ML and evals.

Portfolio at g26karthik.dev LinkedIn Email

HackerRank Orchestrate October 2026: 1st of 4,600 Amazon ML Challenge 2026: 38th of 32,000+ teams
28 open-source contributions: 25 merged pull requests and 3 credited fixes Three first-author papers Langfuse Security Hall of Fame

Hi, I'm Karthik. I build AI systems and make sure they still work on data they have never seen: agents and retrieval that I test properly, LLM quantization, quantum ML, and evaluation work where the hypothesis is written down before the experiment runs.

I also like fixing things in other people's code: 28 open-source contributions so far, including PyTorch, pandas and mem0. I'm in the final year of my B.Tech in CSE (Geethanjali College of Engineering and Technology, Hyderabad) and looking for applied AI and ML roles.

Featured projects

REFEREE: claim verification for research papers
built at RamanIQ multi-agent Docker

  • Reads a paper, lists its main claims, runs a test for each in Docker and writes a reviewer page: Verified or Not verified, with the evidence.
  • ~87% accuracy on evals and tests, 185 kernel tests, and 36K+ papers screened to curate 380+ reproduction candidates.

AlgoSentinel: autonomous PR auditor
agents Docker sandbox Gemini

  • Profiles the old and new version of every changed function in a sandbox and fits growth curves to catch O(n) to O(n²) regressions that tests miss.
  • 55 tools across 4 namespaces, subagent orchestration, 26 tests, and an eval harness at F1 1.00 on 5 golden cases.

FounderOS: voice-first AI chief of staff
🥈 2nd prize Qdrant parallel agents · live demo

  • Turns spoken thoughts into decisions, tasks and ideas with 5 Lyzr agents (three extractors run in parallel) over 4 Qdrant memory collections.
  • Live in production with a privacy-first design: raw conversation stays private.

Buy or Wait?: cash-flow decision engine
exact Decimal LLM perception independent verifier

  • Recommends pay, wait, installments or decline from a 90-day forecast. The LLM reads documents; exact-arithmetic code does the money.
  • 318 tests across 18 suites, with 233 of 250 decisions identical to an independent reference evaluator.

Closer: deadline agent on Google ADK
planner evaluator loop human-in-the-loop

  • Turns a goal and deadline into finished work: draft emails, calendar blocks, research and documents, pausing only for real decisions.
  • 58 tests, a judge that grades each specialist's output with an automatic retry, and a scheduled monitor for closing deadlines.

BizMatch: Amazon ML Challenge 2026
🏆 38th of 32,000+ entity resolution team lead

  • Finds every real copy of a business across ~10M noisy records and rejects look-alike decoys: e5 retrieval, a LightGBM cascade and cross-encoders.
  • 0.921 to 0.990 F0.5 on free compute. Engineering write-up.

More: ParallaxOps (3D reconstruction platform) · Nishchay (COD risk engine) · WhatsApp Notification Router · Grounded Support Triage · Setu (district health network) · EdgeSentinel · Avia · full list on the portfolio

Research

First author on all three.

  • Geometry conservation in variational quantum classifiers (QNC): Quantum Machine Intelligence (Springer), major revision · paper and code
  • Language-aware calibration and bit allocation for low-bit quantization of Indic language models: RAIT 2027, under review
  • CANARYCODE: sensitive-data leakage in LLM coding agents: DAI 2026, under revision

Related open projects: KVQuant Lab and TraceWarrant. More on the portfolio.

Experience

Role Where What I did
Applied AI Engineer Intern RamanIQ (AI startup) · Aug – Oct 2026 REFEREE, an autonomous research-review product: multi-agent orchestration plus deterministic pipelines for paper analysis, experiment execution and evidence-backed verdicts, at ~87% accuracy on evals and tests. Screened 36K+ papers to curate 380+ reproduction candidates. 19+ PRs shipping connectors, UI and backend features
Tech Engineer, AI & Backend Hecta Proptech · Jun – Jul 2026 CatBoost + GPU embeddings over 175K+ auction notices: 84.3% ROC-AUC, 80% less manual review. A Cox survival model (0.79 C-index) behind a validated Flask API, and the first hybrid RAG platform. Features serving 5K+ users a day
Full Stack Engineer (sole) chkchk (stealth startup) · Jan – Apr 2026 A creative-talent marketplace end to end on Next.js and Supabase: −50% DB calls, −30% latency, realtime, EC2 + CI/CD, and the Android app. 66 of 74 commits
AI & Frontend Development Intern Edunet Foundation · Aug – Oct 2025 A lecture audio-to-notes generator (Whisper, T5, DistilBART) for 60+ minute lectures; GPU-accelerated Whisper with caching cut processing time by ~80%
Web Development & SEO Intern Badakarobar · Mar – Jun 2025 Responsive landing pages with ~45% better Lighthouse scores, SEO and metadata work, and Python-generated sitemaps

Open source

I started with small fixes and moved to the bigger libraries I use every day. In total that is 28 contributions: 25 merged pull requests across 19 projects, plus 3 fixes that maintainers landed themselves and credited me for (Langfuse, PyTensor and PyMC). Langfuse also lists me in its Security Hall of Fame. See the merged PRs (PyTorch lands PRs through its merge bot, so GitHub's filter shows one fewer.)

Project What I contributed Status
PyTorch · 105K★ SGD now rejects NaN lr, momentum and weight_decay instead of silently training on NaN, matching the other optimizers ✅ Merged
mem0 · 68K★ The Azure AI Search vector store for the TypeScript SDK: hybrid search and quantization, 754 lines ✅ Merged
Hindsight · 58K★ Fixed audit-log rows silently lost to garbage-collected write tasks and at shutdown ✅ Merged
pandas · 50K★ A PyArrow datetime-merge regression test, Ruff B905 enforcement, and a docstring migration ✅ 3 Merged
Langfuse · 36K★ Found a prompt-cache key collision (version 3 vs label "3" served each other's prompt) 🚀 Shipped via #15040
matplotlib · 23K★ Documented the TwoSlopeNorm linear-colorbar workaround ✅ Merged
sktime · 10K★ Fixed broken pytorch-forecasting references in the forecaster adapters ✅ Merged
SQLFluff · 10K★ A PostgreSQL OPERATOR(schema.op) parser, 649 lines with fixtures ✅ Merged
HackerRank hiring-agent · 7.3K★ Fixed a PyMuPDF memory leak and a Windows UTF-8 crash ✅ 2 Merged
authlib · 5.4K★ OAuth endpoints now answer insecure-transport requests with a proper error response instead of raising ✅ Merged
PyTorch-Ignite · 4.8K★ ReduceLROnPlateauScheduler now honours min_lr for every param group, not only the first ✅ Merged
NumPyro · 2.8K★ Hurdle distributions over positive bases now include zero in their support, and init_to_mean falls back to the median when the mean is not finite ✅ 2 Merged
QuTiP · 2.1K★ rand_dm and rand_herm no longer hang forever when all eigenvalues are equal ✅ Merged
sbi · 882★ Corrected the mixture-density-network log_prob under structured z-scoring, and Sinkhorn now checks convergence on the marginals and warns when it fails ✅ 2 Merged
PyMC ecosystem · pymc-marketing, arviz-stats, pymc-extras Moved pymc-marketing off a deprecated dataset call and fixed its notebook equations, kept the two largest integers in separate arviz-stats histogram bins, and traced a Windows-only Pathfinder test failure in pymc-extras to numba cache numerics ✅ 4 Merged
PyMC · PyTensor Diagnosed the wrong ExGaussian and SkewStudentT means and a Numba Cholesky compile failure on read-only inputs; the maintainers landed both fixes with me as co-author ✅ 2 Co-authored
qutip-qip · toqito Processor.run_state no longer mutates the caller's collapse operators, and kraus_to_choi accepts wrapped completely-positive Kraus forms ✅ 2 Merged

Tech stack

Languages, frameworks and infrastructure

LLMs & ML: OpenAI · Claude · Gemini · LangGraph · Hugging Face · LightGBM · CatBoost · XGBoost · scikit-learn · FAISS · BM25 · GPTQ · PennyLane

Pinned Loading

  1. grounded-support-triage grounded-support-triage Public

    LangGraph support triage across HackerRank, Claude, and Visa corpora. Hybrid BM25+dense RAG, citation critic, eval CLI, production observability.

    Python 3

  2. tracewarrant tracewarrant Public

    Content-free telemetry conformance and simulator-neutral validation of performance predictions for agentic systems.

    Python 1

  3. algosentinel algosentinel Public

    Autonomous PR agent that profiles changed functions in Docker to catch O(n)→O(n²) regressions. Gemini orchestration, 55 tools, FastAPI, eval harness.

    Python

  4. FounderOS FounderOS Public

    Voice-first AI Chief of Staff for solo founders

    JavaScript

  5. whatsapp-notification-router whatsapp-notification-router Public

    Personalized notify / digest / mute routing for multimodal WhatsApp messages. Rank 20 of 1,983 at HackerRank Orchestrate (Aug 2026), with an honest post-mortem.

    Python

  6. ghost-data-still-exists ghost-data-still-exists Public

    An audit agent that proves whether a person's data is really gone: resolves identity across systems, judges state at the audit cutoff, finds surviving PII and restore paths, and plans the fix. Firs…

    Python 22 4