Amit Saharana
#3 in India on Codeforces (chintu) · Founding AI Systems Engineer
IIT Delhi CS · Ex-Facebook · Ex-100ms (E5) · CTO Wokay.
Building PlanCodeShip — on-device AI apps for India. Fine-tuned LLMs, llama.cpp on ARM64, Hindi voice input via Shruti STT.
Background
About Me
I'm a systems engineer with 10+ years scaling low-latency distributed infrastructure — from Facebook's crash reporting to 100ms's multi-region media routing engine with AI transcription to Wokay (CTO) where I architected the LLM orchestration engine, multi-model routing, and enterprise RAG pipeline across 50K+ daily interactions.
Currently building EdgeHub at PlanCodeShip — three offline AI apps (government schemes, crop advisory, NCERT tutor) running on $100 Android phones. Compiled llama.cpp for ARM64 via NDK under sub-1.5 GB RAM limits, fine-tuned Qwen 2.5 with LoRA on MLX, optimized through F32→F16→Q4_K_M quantization, built a JNI bridge with UTF-8 Devanagari handling, bundled SQLite with FTS5, and integrated Shruti STT (~29 MB) for Hindi voice input. ~22.5 tok/s on ARM64.
PlanCodeShip is the lab. ChatGo is the commercial product (WhatsApp Business cost optimization — 6 tools, up to 75% savings). Jev is the safety layer (agent tool-call guardrails, under 200ms). AutoSquad is the agent platform (multi-agent orchestration with cascading model routing at 85%+ blended margin).
Indian Institute of Technology, Delhi
Attended B.Tech in Computer Science and Engineering · 2011 – 2015
Undergraduate Research (SURA): "Algorithms for High Dimensional Proximity Problems" under Prof. Sandeep Sen
AIR 79 in AIEEE 2011 (out of 1.15M candidates)
KVPY Fellow (2009) — National science fellowship for exceptional aptitude
NTSE Scholar (2007) — NCERT national scholarship
Core Systems & Languages
AI Systems & Inference
Infrastructure & Data
Recognition
Achievements
Codeforces Grandmaster — #3 in India
Peak rating 2416 (handle: chintu). Ranked #3 in India. Top-10 in India across Codeforces, TopCoder, CodeChef, and HackerRank.
Google Distributed Code Jam — #1 in India
61st globally in Google's Distributed Code Jam 2017. Ranked 1st among all Indian participants.
Hackathon Wins
- 1st Prize — Microsoft Code.Fun.Do 2013 · Inter-college hackathon
- 2nd — Sears Dots & Arrows 2016 · ₹2 Lakh prize
- 3rd — Govt. of India 12th Five Year Plan · National hackathon
- 8th — Zenefits Hackathon 2015 · San Francisco onsite
Portfolio
Projects
AI-native products — multi-agent systems, edge inference, safety guardrails, and cost optimization tools.
Career
Experience
PlanCodeShip
· Founding AI Systems Engineer- Architected an offline-first edge AI inference runtime on ARM64 Android, deploying a fine-tuned Qwen 1.5B model natively via llama.cpp under strict sub-1.5 GB RAM and thermal throttling limits.
- Engineered a custom JNI bridge with UTF-8 Devanagari token handling and cross-compiled llama.cpp via NDK/CMake. Achieved ~22.5 tok/s generation throughput on ARM64 test hardware (Android API 35) across an optimized F32→F16→Q4_K_M quantization pipeline.
- Replaced Android's bundled SQLite with a CMake-compiled SQLite 3.45 with FTS5 enabled, implementing BM25-ranked retrieval and Hinglish synonym expansion with zero external network dependencies.
- Integrated offline Shruti STT (~29 MB) for Hindi/Hinglish voice input, building a deterministic normalization layer (Devanagari↔Latin transliteration, compound splitting) to align transcribed audio tokens directly with the LLM context window.
Wokay
· CTO- Architected "Woby," an LLM orchestration engine processing 50K+ daily interactions, converting unstructured conversational streams into structured actions via agentic tool-calling across GPT-4o and Claude 3.5 Sonnet.
- Implemented a latency- and cost-aware multi-model routing layer with deterministic fallback policies (frontier models → compact low-latency models → local runtime), cutting inference overhead by ~40% while preserving response quality bounds.
- Built an enterprise RAG pipeline incorporating sliding-window document chunking, semantic vector embeddings, and metadata-filtered hybrid search to enforce citation grounding.
- Led the C++ and React architecture for real-time state synchronization, enabling concurrent multi-tenant chat sync and low-latency voice-agent integration.
100ms
· Software Engineer, E5 (Infrastructure)- Architected a multi-cluster media routing engine in Golang across distributed geographic regions, sustaining sub-200ms media relay latency with HIPAA-compliant data residency guarantees.
- Engineered a high-throughput telemetry pipeline streaming 500K+ events/hour to Segment, Kafka, and webhooks, implementing distributed idempotency checks to guarantee exactly-once delivery semantics.
- Built dynamic worker pool orchestration using KEDA on Kubernetes (via Helm), scaling media processing pods elastically based on live room count, transcription queue depth, and GPU inference load.
- Designed an automated media transcription and summarization pipeline using Whisper for real-time speech-to-text coupled with asynchronous LLM extraction for meeting action items.
Adobe
· Computer Scientist- Worked on Adobe Fresco, the cross-platform flagship drawing app — performance-critical rendering pipeline using React, C++, WebAssembly, and UWP
WeCreateProblems (WeCP)
· Software Engineer- Built ML-powered code evaluation platform — automated scoring using AST analysis, code similarity detection, and NLP-based feedback generation across 30+ programming languages
- Designed containerized code execution sandbox on EKS Kubernetes with GitOps (ArgoCD, Terraform, Consul) — isolated runtime environments for secure multi-tenant code assessment
- Implemented decision automation workflows using Camunda with ML-driven candidate scoring pipelines — feature extraction from code submissions, automated grading, and anomaly detection for plagiarism
HackerRank
· Software Developer- Developed code plagiarism detection system using ML-based code similarity — AST fingerprinting, token sequence analysis, and feature engineering for automated cheating detection at scale
- Built sandboxed code execution platform on AWS — ephemeral VMs with state-diff grading to validate submissions by analyzing filesystem and process changes
- Designed microservices for the coding assessment platform in RoR, C++, and CoffeeScript
- Improved crash reporting framework — built categorization pipelines using Hive and Python for automated crash clustering and prioritization
- Optimized crash deduplication algorithms and integrated the framework into the iOS app with performance benchmarks
Let's connect
Get in Touch
Whether you want to discuss AI infrastructure, LLM systems at scale, agent architecture, competitive programming, or just say hi.