Surrey, British Columbia · Tuesday, August 4, 2026 Portfolio Edition
Résumé (PDF)
Production AI
Real-time inference
Distributed systems

Ishan Sharma

Senior Software Engineer
Production AI, on-prem & cloud
Writes on sustainable AI systems
The Lead

He builds AI systems for rooms where being wrong is expensive

The model is the smallest part of the system — he builds the rest: sub-500ms inference at two hundred concurrent sessions, redaction that never leaves the building, and fallbacks designed before they are needed.

By Ishan Sharma·Senior Software Development Engineer, Yactraq Online

M y work sits at the intersection of real-time inference, distributed systems and decision-critical AI — environments where models must perform under uncertainty and tight operational limits, and where an error is not an interesting failure case but a cost someone absorbs.

That framing changes what architecture means. Safe degradation, recovery paths and long-term maintainability stop being nice-to-haves added after launch and become the first questions on the whiteboard. Confidence-sensitive behaviour when signal is low. Fallbacks that were designed rather than discovered during an incident.

It also means instrumentation deep enough to know why a GPU is slowing — KV-cache growth, memory fragmentation, PCIe pressure — long before the symptom reaches a customer. On two H100s I held inference under 500 milliseconds at more than two hundred concurrent sessions by tuning batching and concurrency against numbers I could actually see.

Across all of it runs one conviction: the model is the smallest part of the system. What earns trust in production is everything built around it — deterministic execution for anything that carries a number, curated context over clever prompting, provenance a user can click, fallbacks designed before they are needed. Done right, small bounded models running on hardware you control do work people can bet on.

That is also how to read this site. The case studies are the record — production systems with the constraints and failure stories left in, the same work my résumé compresses to a page. The writing is the reasoning — what operating them taught me about evaluation, drift, ownership and the seams where AI breaks, published as essays on LinkedIn. The résumé says what shipped. The essays say how I think. The case studies are where the two meet.

Selected Work

Three case studies
Case Study 01 · Retrieval

Four retrieval signals, one answer, half a second

A real-time agent-assist engine that turns noisy live-call transcripts into knowledge-base answers while the customer is still talking. A small language model first decides whether the turn deserves a search at all — chatter and filler get an explicit no-query verdict rather than a bad answer — then compiles the messy speech into a structured query with product and concept hints validated against a closed vocabulary.

Retrieval fuses four disagreeing signals: lexical match and dense embeddings combined by reciprocal-rank fusion, then a cross-encoder pass, then an instruction-tuned reranker that scores each document against the actual support task rather than generic similarity. Absolute and relative score floors sit under all of it, session memory carries confirmed topic keywords across turns and decays them the moment the conversation shifts, and account numbers are redacted before any model sees a word. When the floors are not met, the agent simply is not interrupted.

Behind it, a thousand-document knowledge base — web content, PDFs, spreadsheets, image-only scans read by a vision model — is distilled into scored, agent-facing cards. The latency budget taught the real lesson: the hard part was never any single retriever. It was making four signals that disagree resolve into one defensible suggestion, in half a second, on hardware the customer controls.

More detail
LLM query formulation with no-query gating over live ASR transcripts; session keyword carryover with topic-shift decay; Elasticsearch-native BM25 + kNN with reciprocal-rank fusion; cross-encoder plus instruction-tuned reranker cascade with logit-calibrated score spreading; absolute and relative score floors; number redaction before any LLM call; per-session document dedup and trustworthiness scoring; ingestion pipeline distilling web content, PDFs, spreadsheets and VLM-read image scans into usefulness-scored agent-facing markdown cards.
Python · Elasticsearch · e5 embeddings · BGE reranker · vLLM · On-prem GPUs
Lexical · BM25Dense vectorsCross-encoderInstruction rerankRRF fusion& score floorsConfidencegateAnswerAbstain< 500 ms end-to-end

The four-signal fusion path, with confidence gating before generation.

Case Study 02 · Audio

The customers we could not reach, and the capture layer that changed it

MicrophoneSpeaker audioDesktopcapture layerStreamASRbackendsNo SIP stack required

Contact centres without a SIP telephony stack were simply unreachable. I built a framework that reliably streams both microphone and speaker audio from the desktop to our ASR backends — turning an integration dead end into a competitive edge and opening deployable segments the product previously had to walk away from.

More detail
Real-time dual-channel capture on the Windows audio stack: microphone and speaker audio taken directly from the device layer, so speaker attribution comes from physical topology — agent on the microphone, customer on the sound card — with no diarization model in the path and none of its error rate. Silence is signalled explicitly on each channel rather than left for the transcription backend to infer from absence; audio is buffered and shipped over a message queue to streaming ASR. Built as a desktop application (Electron + React); opened contact-center deployments where SIP or telephony APIs cannot be exposed, and remains an ongoing production capability rather than a one-off integration.
Case Study 03 · Agents

An agent that plans its own retrieval, then asks a human to check

PlanRetrieveMongoDBSearchCRMReportHumancheckCurated plans, reused

A supervisor-facing analytics agent that answers plain-language questions over MongoDB, Elasticsearch and CRM data. One LLM planning call is compiled into a validated, deterministic execution DAG, so every number in the answer is typed, derived and traceable — the model never words a figure it did not receive. Interpretation is surfaced as user-reversible filter chips, and human-curated plans are stored and reused, so retrieval flows keep improving instead of quietly drifting.

More detail
One planning call on a locally hosted quantized model produces a typed execution plan; deterministic code validates it against the live schema, repairs it within strict bounds, executes it step by step, and templates every trust-bearing sentence — scope, counts, sample basis — so the model never words a number it did not receive. Counts are derived or omitted, never estimated; evidence is an exact, clickable set of source records. Question routing sends why-questions down a fully deterministic path with no generated queries at all. Reinterpretations surface as removable filter chips that re-execute the cached plan, so the stochastic planner is never consulted twice about the same question. Guarded by a golden-test suite and a regression corpus harvested from real production plans; shipped as a React component embedded in the existing Angular dashboard over a streaming connection.

In Brief

Shipped & in progress

On-prem PII redaction

PCI-oriented NER with 50+ configurable labels, redacting in real time inside the customer's own network.

Mongo → Elastic pipeline

Ruby indexing supporting safe reindexing and schema evolution under live production traffic.

Decentralized container execution

Orchestration for containerized workloads on GPU-backed DePIN networks, as a cloud alternative.

Alzheimer caregiving assistant In development

Builds a family graph from photos via face clustering, then retrieves context to help caregivers recall people and shared memories.

AI Confidant In development

Real-time speech-to-speech support for NGOs — streaming ASR, LLM dialogue, TTS and safety filters under high stress.

Zero — reflection assistant In development

Private journaling for spoken and written input, with low-latency ASR and encrypted capsule memory.

The Chronicle

2021 — Present
2025 →
Senior Software Development Engineer

Own production agent-assist systems for contact centres: on-prem LLM summarization, real-time PII redaction, and a Ruby indexing pipeline reindexing under live traffic. Delivered sub-500ms inference at 200+ concurrent sessions on 2×H100 by tuning batching and concurrency, and built GPU instrumentation to catch degradation before it lands.

More detail
Inference is tiered by latency class: H100s where the deadline is half a second, a single consumer GPU where it is thirty minutes — the batch tier clears 3,500+ transcripts a day on one card, with VRAM budgeted across co-tenant workloads. Hardened that pipeline from a 30% error rate to zero by building its observability — journey logs, metrics, alerts — from scratch, and cut agent-assist evaluation latency from 7.5s to 1s with right-sized models. Model upgrades are gated by golden-set evaluation against human QA. Also ships a supervisor-facing analytics agent: one planning call compiled to a validated, deterministic execution plan, delivered as a React component inside the existing Angular dashboard. 10+ on-prem enterprise go-lives in 18 months, engineering live within a day.
2025 →
Founding Software Engineer

Led technical design of a decentralized container execution platform running workloads on GPU-backed DePIN networks. Defined execution flows, isolation boundaries and deployment primitives, and built orchestration to schedule and monitor across heterogeneous GPU environments.

More detail
MaxCore makes fragmented compute markets usable for reliable AI workloads: an opportunistic execution layer that evaluates available infrastructure against user-defined preferences — cost, geography, performance, reliability — and places each workload on the best option across containers, VMs and serverless, without provider lock-in. He designs the orchestration and resilience model: when low-cost capacity disappears, workloads fail over to a ready alternative while a replacement is provisioned in parallel, then rebalance back as preferred capacity returns. The work spans provider abstraction, workload portability, placement strategy, failover, recovery and state reconciliation — plus the Next.js platform, the React company site, and a 20-page GPU-infrastructure RFP with compliance matrix and bill of quantities.
2024
Information Technology Developer

Built citizen-facing Angular features including “Authorize My Representative,” and worked across Angular and Java services on the “File a Formal Dispute” workflow — accessibility, responsive behaviour and regression coverage for services taxpayers depend on.

More detail
Scoped business requirements directly with stakeholders, then implemented them end to end across the Angular front end and Java services.
2022 – 2024
Full Stack Software Engineer

Built real-time ASR pipelines on Kaldi and NVIDIA NeMo with high-availability architecture on AWS and distroless Docker images. Led the Angular 14 migration of the BI dashboard, and shipped the PII redaction system, a conversational SQL query assistant, and the desktop audio capture framework.

More detail
Multilingual ASR pipelines with Python and MongoDB services behind them, run as highly available speech-processing deployments on AWS. The capture layer reads speaker attribution from physical device topology — agent on the microphone, customer on the sound card — so no diarization model sits in the path, and silence is signalled explicitly per channel rather than inferred from absence; roughly a third of customers run on this path today. The role grew from machine-learning and backend work into ownership across architecture, infrastructure and customer-facing applications — the foundation for leading the real-time Agent Assist and GenAI initiatives that followed.
2021
Data Engineer
Datafact.io · Bangalore

Server-side clickstream tracking for Shopify merchants on GCP with Tag Manager, Pub/Sub and containerized ETL — multi-tenant behavioural collection, plus async pipelines and CI/CD that cut manual setup per client.

Opinion

Syndicated from LinkedIn →

Correspondence

Let's find twenty minutes.

The conversations I enjoy most are about production AI systems that have to be fast, private and dependable — if you're wrestling with one, my inbox is open.

Common questions
Has he handled production incidents?
Yes. When a customer reported unexplained data leaving their on-prem environment with his employer's software as the suspect, he worked logs, packages and service health in front of the customer's security team, isolated the cause to a crash-looping legacy service repeatedly pulling packages from an external host, reported it straight to the customer, and decommissioned the service. He has also restored a stalled production pipeline in about thirty minutes — a regression from his own team's update, which he says plainly — by finding a child process whose output pipes were never drained.
What is distinctive about how he builds?
He looks for the version of the problem that does not need a model: speaker attribution read from device topology instead of a diarization model; a deterministic retrieval path instead of a generated query; a single planning call compiled to a validated execution plan instead of an agent loop. Models do the work only where they earn it — everything that carries a number is deterministic code.
Who is Ishan Sharma?
Ishan Sharma is a Senior Software Engineer based in Surrey, British Columbia, Canada. He builds production AI systems — real-time LLM inference, retrieval-augmented agents, speech recognition pipelines, and on-prem PII redaction — built to be fast, private and dependable.
What does Ishan Sharma work on?
He currently owns production agent-assist systems for contact centres at Yactraq Online: on-prem LLM summarization, real-time PII redaction with 50+ configurable labels, and retrieval pipelines that hold inference under 500 milliseconds at 200+ concurrent sessions on two NVIDIA H100 GPUs.
How can I reach Ishan Sharma?
Email (ishansharma1320@gmail.com) or LinkedIn — he reads both. The conversations he enjoys most are about production AI systems that need to be fast, private and dependable.