LLM systems that reconcile, cite their source, and fail safely.
I take LLM prototypes into production for regulated finance, compliance, and enterprise operations. The demo is the easy part. Holding up under load and under audit is the work.
Track record
NIST 800-53Control logic shipped into retrieval and prompting
The gap between a demo and a system
If two or more of these are true, we should talk. Every one is something I have personally fixed in production.
-
Your extraction pipeline works on 10 sample PDFs and falls apart on the real corpus.
Document Intelligence Pipeline → -
The model returns null and nobody knows whether it was a content filter or a token limit.
Production LLM Triage → -
Nobody can tell you what the system costs per document.
LLM Production Readiness Review → -
Legal and audit are blocking the launch, and there is no access model or audit trail.
LLM Production Readiness Review → -
The demo hit 90% and production is at 60%, with no eval harness to explain why.
Production LLM Triage →
What I get hired to fix
Narrow on purpose. This is the work where being wrong is expensive: extraction feeding financial reconciliation, retrieval answering regulatory questions, AI services that legal has to sign off on.
Four core capabilities
Document intelligence & extraction
PDF and scan to structured, validated data. Chunking strategy, sliding-window context, token budgeting, and an accuracy target agreed before work starts.
Reliability & guardrails
Model fallback cascades, explicit failure recovery, real-time PII masking, prompt-injection and toxicity defense, and an eval harness that catches regressions.
Audit-grade AI architecture
Enterprise SSO, hierarchical RBAC, actor attribution, structured audit logging, and approval workflows that satisfy an internal audit team.
Cost & performance
Model routing, token budgeting, inference throughput tuning, and a per-document cost model you can put in front of a CFO.
End to end on every one of them
-
01
Research
Feasibility spikes, model and approach selection, and targeted work like the semantic alignment that took compliance-mapping F1 from 30 to over 70. Delivered inside client engagements, on their problem.
-
02
Architecture & design
Service boundaries, API contracts, data schemas, and the access model, all designed to survive an internal audit long before one gets scheduled.
-
03
Build & debug
I write the code, and I find out why yours returns null. Python, FastAPI, React, and the failure modes underneath them.
-
04
Deploy & DevOps
Kubernetes, Helm, Docker, CI/CD, and observability. You get a running system and the pipeline that keeps it running.
-
05
Test automation
Eval harnesses, version-controlled eval sets, regression gates in CI, and load tests that find the limit before your users do.
-
06
Handover
Runbooks, architecture decision records, and a team that can operate the system without me on the call.
Shipped, under real constraints
Anonymized and cleared of client-identifying detail — no platform names, no topologies, no internal identifiers.
Scaling a regulated-finance RAG copilot to 10,000+ users
Classification and synthetic-data APIs cut task turnaround by half.
- Led design and delivery of the copilot end to end, then scaled it to 10,000+ users through combined infrastructure and algorithmic work.
- Rebuilt the retrieval stack in four separately measurable stages: parsing, query extraction, ranking, generation. Tuning the prompt and hoping is not a debugging strategy.
- Targeted research on semantic alignment for regulatory compliance mapping, the single highest-leverage accuracy problem on the platform.
Three more pieces of the build
- Redesigned the ML inference pipeline for throughput under scaled load.
- Deployed and fine-tuned open-source LLMs in house for generation and similarity, cutting external API dependency and per-request cost.
- Established DVC and MLflow pipelines so data versions and experiments were tracked, which made every accuracy gain attributable to a specific change.
- Fix retrieval before you touch the model. Most bugs reported as "the LLM is wrong" are ranking bugs wearing a costume.
- Version your eval set like source code. Without it you cannot distinguish a real improvement from run-to-run noise, and you will ship regressions confidently.
- Get the customer to measure accuracy on their own corpus. Your internal number will not survive procurement; theirs closes the deal.
Automating vendor invoice reconciliation at airport scale
- Architected and own end to end a platform that replaces manual invoice review with automated extraction, reconciliation, and multi-stage approval.
- A resilient PDF-to-structured-data pipeline behind an enterprise API gateway: three-model fallback cascade, page-chunked processing with sliding-window context, and token budgeting. Explicit recovery paths for content-filter and truncation failures turn brittle single-shot parsing into a production-reliable one.
- A reconciliation and savings engine with period-aware accounting: potential separated from confirmed savings, cross-period rebooking, multi-currency normalization, discrepancy and overrun detection, and content-hash duplicate control against a cloud data warehouse.
Three more pieces of the build
- Enterprise SSO via OIDC with JWT and JWKS validation behind the gateway, hierarchical scope-based RBAC across local, regional, and global scopes, plus actor attribution and lifecycle audit logging built for internal audit review.
- An executive weekly-status platform producing audit-friendly decks with deterministic text fitting, metric reconciliation, and multi-portfolio rollups.
- A customer-service response system on a dual-summary architecture with layered PII masking, prompt-injection detection, toxicity validation, and strict output-format enforcement.
- Reconcile against a system of record, never against the model. Extraction confidence is not correctness, and the difference becomes visible the first time finance disagrees with your output.
- Handle the two null-content failures separately. A content filter and a truncated response both give you nothing back, but they need completely different recovery paths, and retrying a token-limit failure unchanged just burns spend.
- Decide on day one whether a number is potential or confirmed. Finance will ask, and retrofitting that distinction through a schema and a UI is far more expensive than designing for it.
Making control-mapping answers traceable to the clause
- Designed and led an enterprise RAG platform for compliance and policy automation: context-aware document retrieval, security-control mapping, and LLM-driven remediation recommendations.
- Encoded NIST 800-53 and CMMC control logic directly into the prompting and retrieval strategy, so outputs trace to specific regulatory clauses by construction.
- Exploited framework overlap, where one underlying control satisfies requirements in several frameworks, so mapping effort compounds across FedRAMP and CMMC.
- Put control logic in retrieval and prompting. A post-hoc filter cannot produce traceability, traceability is the whole deliverable, and you cannot add it afterwards without a rebuild.
- Make every answer cite a control ID. Assessors accept citations; they do not accept confidence scores, and neither will your customer's auditor.
- Map controls once, centrally. Framework overlap is where the leverage is: treat FedRAMP, CMMC, and NIST 800-53 as views over one control model instead of three separate projects.
Six ways to bring me in
Pick by what is in your way right now. All six cover the full lifecycle and run on one hourly rate, so what you choose changes the cadence and the commitment while the rate stays the same.
Start here
Production LLM Triage
Accuracy fell off a cliff, the model returns nothing, costs tripled, or nobody can reproduce last week's numbers. I get into the logs and the retrieval stack and tell you what is actually wrong, in days.
LLM Production Readiness Review
A written audit of what you are running: failure-mode inventory, cost-per-document model, guardrail and eval gaps, and a prioritized remediation roadmap. Where most engagements start.
Then build it
Prototype → Production Build
Take a working notebook or demo to a deployed, monitored, guardrailed service. Cloud infrastructure, CI, and observability included, plus an eval harness and load tests that catch regressions before your users do and a runbook your on-call can actually use.
Document Intelligence Pipeline
PDF or scan to validated structured data, with reconciliation against your source of truth. The accuracy target is agreed in writing before work starts, and the pipeline is built to hit it on your real corpus.
Fractional AI Architect
Senior AI architecture without a full-time hire. I sit in your architecture review, vet vendor and model decisions, hold the line on eval discipline, and mentor your ML team.
MVP for a Funding Round
A working system investors can actually use, not a clickable prototype. Built on the same architecture I would ship to a regulated buyer, so it survives the technical diligence call and does not need a rewrite the month after the round closes.
The things you were going to ask anyway
Remote, with daily overlap
Bengaluru-based, working US Eastern hours. Standups, reviews, and incident response happen inside your working day, not after it.
I write things down
Architecture decisions, failure modes, and runbooks are deliverables, not afterthoughts. When the engagement ends, the reasoning stays with your team.
Your internals stay yours
Whatever I publish or say about your system gets the same treatment the case studies above got, and you clear it first. Your codebase, your architecture, and your incidents stay inside the engagement.
Hourly, capped, and logged
You approve an hours cap before work starts and get a weekly log of what those hours went to. Scale the commitment up or down as your priorities move — no change order, no renegotiated statement of work.
What I build with
Grouped by what it is for, because a logo wall tells you nothing.
GenAI & LLM systems
Models & frameworks
Backend & data
Cloud, DevOps & security
Aman Jaiswal
Six years in applied AI/ML, four of them shipping production generative AI. I started as an ML engineer building NLP and computer-vision systems, spent two and a half years scaling regulated-industry RAG to 10,000+ users, and now consult independently on LLM systems that have to survive an audit.
Most enterprise AI projects do not fail on the model. They fail on reliability, eval discipline, access control, and the unglamorous work of making output defensible when someone reports the numbers upward. That is the part I am good at.
Available for remote engagements at 20–40 hours per week.
Education
B.Tech, Computer Science — CGPA 8.79/10. Certifications: LLMOps (Udacity), Introduction to Machine Learning in Production (DeepLearning.AI), Computer Vision Nanodegree (Udacity).
Publications
- Artificial Intelligence Analysis of Cultural Narratives Shaping Emotional Responses to Infertility — IEEE, 2025 · paper
- Earlier work: Analysis of Deep Learning Algorithms on COVID-19 Radiography Database and AUTHEER: A Voice-Based Speaker Authentication System — both IJAST
People who've worked with me
Senior collaborators vouching for the work and how I operate. Reach them directly on LinkedIn.
Aman has contributed a lot to our NLU service, He is well-versed in NLP and ML, has great enthusiasm about the tech, and is always up to solve complex problems.
He is going to achieve great carrier milestones with his thrust for learning.
Aman pairs real research depth with the discipline to put models into production safely. On our work together he's the one who insists on evaluation rigor and governance before scale — exactly what regulated teams need and rarely get.
Common questions
Why one person instead of an agency?
With an agency there is usually a layer between you and whoever writes the code, and the person who scoped the work is rarely the person who does it. Here they are the same person. The honest tradeoff: I am one person's throughput, I take a small number of engagements at a time, and if you need four workstreams running in parallel you want a firm. What you get in exchange is that the feasibility research, the build, the deployment, and the eval discipline sit in one head, so nothing is lost in a handoff — which is where most of these projects actually come apart. When a job needs hands I do not have, I tell you early.
How do I know the hourly billing will not run away?
Every engagement has an hour cap agreed in writing before I start, and I do not bill past it without your sign-off. You get a written log every week: hours spent, what they went to, and what is next. If the work will need more hours than we scoped, you hear it while there is still time to decide rather than in an invoice. It is also why most engagements open with a Readiness Review of 20 to 40 hours — a small, capped commitment that ends in a written deliverable you keep whichever way you go next.
What do you need from us to start?
For a Readiness Review: read access to the repo, a sample of real inputs, and 30 minutes each with whoever owns the model, the data, and the compliance sign-off. For a build: the same, plus a decision-maker who can approve architecture changes without a committee.
Thirty minutes to work out what you actually need
Free, no deck, no pitch. Bring an architecture diagram, a stack of failure logs, or nothing but an idea and a deadline. If something is already running we'll go through what breaks first, and if it isn't we'll work out what is worth building. Either way you leave with three things worth doing, whether or not you hire me.
Prefer email? Tell me, in four or five lines:
- What you're building, or want to build.
- What's in your way — accuracy, cost, reliability, audit sign-off, or not yet knowing which approach is right.
- Your timeline.
- Whether you want triage on something broken, a Readiness Review, a build, an MVP, or fractional advisory.
I reply within 48 hours on weekdays. I take a small number of engagements at a time.
Get in touch
Book the call, or email me directly.