ashish.uike available

Software engineer,
ML systems & backend infra

Three years building production ML pipelines and distributed services at HCLTech. I work on the unglamorous half — retries, batch fault tolerance, grounding gates, latency budgets — the parts that decide whether a system survives contact with real traffic. IIT Indore, mechanical engineering. Open to SDE and MLE roles.

open to work Hyderabad, IN Python · C++ · SQL 3 yrs production
trace: career.hcltech & side-builds 2023202420252026
▸hcltech.engagement
Oct 2023 – present · SDE / Technical Lead
role=individual_contributor engagements=2 ownership=end_to_end
  • Individual contributor with end-to-end technical ownership of ML and automation workstreams across two client engagements.
f1=0.91 auc_roc=0.94 batch=monthly sink=oracledb
  • Owned the full lifecycle as sole engineer on the ML workstream — data pipeline architecture, model selection, deployment, monitoring — reporting directly to stakeholders.
  • Engineered a resilient pipeline processing 4M+ records monthly across distributed systems, with fault-tolerant batch handling, automated retry, and validated output persistence to OracleDB.
  • Designed incremental learning so models update continuously without full retraining; ensembles held F1 0.91 and AUC-ROC 0.94 as telecom billing patterns drifted.
  • Tuned Random Forest and CatBoost ensembles with SHAP explainability, catching billing discrepancies before customer-facing release to cut escalations.
agents=7 endpoints=18 tests=50+ retrieval=faiss×2
  • Designed a 7-agent orchestration architecture — understanding, knowledge retrieval, similar-ticket retrieval, resolution, validation, human approval, knowledge generation — as an async stateful graph in LangGraph, isolating each reasoning step into its own failure domain and exposing the pipeline over 18 FastAPI endpoints.
  • Applied fault-tolerance to every LLM-backed call: retry with backoff, strict JSON-schema validation, graceful degradation to safe defaults on timeout or malformed output, each agent held to a sub-5-second latency budget.
  • Built a human-in-the-loop approve/modify/reject gate with immutable audit logging on a 7-table SQLite schema, keeping every automated resolution traceable and reversible.
  • Implemented dual FAISS indices (TF-IDF + SVD) plus a confidence-scored grounding gate that checks each resolution against retrieved evidence before a reviewer sees it.
  • Verified with a 50+ test Pytest suite across a 6-page Streamlit review UI, behind a provider-agnostic adapter layer extensible to ServiceNow, Jira and email.
institution=indian_institute_of_technology_indore
  • B.Tech in Mechanical Engineering, CGPA 7.6/10. Moved into software through algorithms and systems work alongside the degree.
side builds — shipped, tested, public

llm-reliability

Aug – Sept 2026 · published to PyPI

Local-first root-cause engine for LLM, RAG and agent systems. Ingests native, OpenTelemetry and LangSmith traces, emits evidence-backed hypotheses with confidence scores across 9 failure categories — retrieval failure, hallucination, agent loops, tool errors, schema violations. Ships a zero-overhead async tracing SDK, a streaming latency profiler (TTFT, tokens/sec, stall detection), and a multi-run regression comparator tracking p50–p99 drift.

267tests
90%coverage
45labeled scenarios
Python · Pydantic v2 · OpenTelemetry · pytest · mypy strict · CLI
pypisource

advanced-rag-system

Sept – Oct 2025

Document Q&A platform built to production shape rather than notebook shape. Hybrid retrieval over semantic and BM25, multi-format ingestion (PDF, DOCX, TXT, CSV, MD) into ChromaDB, and a Redis cache layer with intelligent invalidation that cut repeat queries from 1.308s to 0.011s.

117×cache speedup
60%hit rate
100%suite passing
FastAPI · ChromaDB · Redis · Docker · LangChain · HuggingFace
source

distributed-url-shortener

Dec 2025 – Jan 2026

A sharding and observability exercise taken to completion. Application-level sharding across three PostgreSQL nodes, multi-layer Redis caching, event-driven analytics through Redpanda, and full Prometheus/Grafana instrumentation across a 13-container compose stack.

1,094req/s
2.7msp50
13containers
FastAPI · PostgreSQL · Redis · Redpanda · Prometheus · Grafana
source
stack — what I reach for

languages

Python · C++ · SQL

core

Data structures & algorithms · System design · Distributed systems · Concurrency · DBMS · Operating systems

backend & distributed

FastAPI · PostgreSQL · Redis · Microservices · Consistent hashing · Horizontal scale-out · Event-driven design · Database sharding · CI/CD · REST · OpenAPI

infrastructure

Docker · Redpanda (Kafka) · Prometheus · Grafana

ml & data

Random Forest · CatBoost · SHAP · RAG systems · LangChain · LangGraph · FAISS · HuggingFace · ChromaDB · Pandas · NumPy · OpenTelemetry · Streamlit

certification

GitHub Copilot, Microsoft
Feb 2026 – Feb 2028
ID 27EB2EB9F3468542

signals — algorithms practice
LeetCoderating 1601 · 650+ solved1601
InterviewBitinstitute rank#13
GeeksforGeeksinstitute rank#21

Let's talk about the hard parts

Backend, ML infrastructure, retrieval systems, anything that has to stay up. I read everything that comes in.

ashishuike8@gmail.com +91 72475 80704 linkedin github
built by hand · Hyderabad, India · 2026