M.
click anywhere to skip
AI/ML Engineer·Remote / Delhi, India·v2.4.1
$ cat ./manifesto.txt

Automatingtheboring.
Scalingtheinteresting.

→Agentic Workflows.—built for your team.

AI/ML engineer building multi-agent stacks and clever automation.
Writes code that nags less than humans.

// python · llms · n8n · postgres · production-grade · under maintenance
manveen@prod:~/portfolio$
book a 30-min audit→see the work ↓email →
taking 2 projects this quarter·build & retainer·replies within 24h
~/engineer.pylive
# manveen.singh — sys.engineer
class Manveen:
    role     = "AI/ML Engineer"
    location = "Delhi · Remote"
    status   = "shipping"
    focus    = [
        "multi-agent systems",
        "workflow automation",
        "production tooling",
    ]
    hates    = ["manual workflows", "unclear logs"]

    def philosophy(self):
        return "If I repeat it twice, I automate it."
python 3.12 · 23 lines✓ no warnings
multi-agentworkflowGTMLLMvector searchembeddingsCVpythonnext.js
// 01system.profile — what i do

Built
for your team.

available · build & retainer
›If your team is doing it twice, I’ll build the system that does it.

I’m an automation engineer who designs, ships, and operates the workflow systems your team is tired of building in-house. Lead pipelines, document processing, multi-agent stacks, internal tooling — built to run unattended, with logs you can actually read.

I don’t hand you a notebook and disappear. I ship production code, wire up the alerting, document the failure modes, and stay on retainer to keep it running. Boring stability is the deliverable.

GTM Pipelines
Lead discovery, enrichment, scoring, and CRM sync — replaces the manual research your SDRs are doing today.
AI Agents & Docs
Resume parsing, document extraction, multi-step LLM workflows — agents that retry, escalate, and report cleanly.
Build + Operate
I don't disappear after shipping. Retainer keeps the system patched, monitored, and on-call when it matters.
No Black Boxes
You'll know what runs and why.
Production-First
Built for uptime, not demos.
Tight Loops
Ship in days, iterate from logs.
Boring Reliability
The best system is invisible.
// 02pipeline.dag — how the systems run5 nodes hot
$ render --graph pipeline.dag --realtime10 nodes · 12 edges
[T-01]TRIGGER[C-02]COLLECT[C-03]SCRAPE[N-04]NORMALIZE[E-05]ENRICH[V-06]VALIDATE[S-07]RANK[R-08]ROUTE[R-09]SCHEDULE[L-10]LEDGER
triggerworkeragentsinkthroughput: 142 evt/s
tail -f pipeline.log
$
// 03selected.work — shipped systems

Systems
already in production.

Each system below is something I've shipped for a team. Stack, outcome, and what it replaced — pick the one that looks like your problem.
OPS-001Lead engineer — architecture · AI · automationshipped
AI-Powered Recruitment Automation

Modular AI recruitment system: NLP resume parsing, embeddings-based ranking over a vector DB, and an automation layer for shortlists and interview scheduling.

challenge:Recruiters were spending 5–10 hours per week manually screening resumes, leading to slow time-to-shortlist, inconsistent evaluation, and recruiter fatigue.
solution:Built a modular system with a Resume Ingestion Pipeline (NLP parsing), an AI Ranking Engine (embeddings + vector DB scoring candidates), and an Automation Layer to auto-generate shortlist reports and schedule interviews.
PythonFastAPIspaCy / LLMEmbeddings + Vector DBPostgreSQLCelerySlack APIDocker
10h → 3h
screening / week
3×
candidate throughput
5d → 1.5d
time-to-shortlist
↑
ranking consistency
OPS-002Architect & buildershipped
PresentAI — Automated Presentation Agent

Text-to-slide agent: structured slide flow, speaker-notes generation, and automated PPT export. Founders go from idea to first draft in under a minute.

challenge:Founders and PMs waste hours structuring slide decks, writing content, formatting layouts, and rewriting messy meeting notes.
solution:Built an automated presentation agent that executes text-to-slide conversion. Implemented structured slide flow, speaker-notes generation, and automated PPT export based on LLM abstractions.
Next.jsTailwind CSSPrismaZustandPostgreSQLGen AIpython-pptxReact DnDClerk
2h → 15m
deck creation
<60s
first draft
↑
pitch iteration
OPS-003GTM Automation Engineershipped
Lead Enrichment & GTM Automation

Multi-source lead ingestion and an API enrichment layer. CRM sync, Slack notifications, lead scoring, and automated data validation.

challenge:Lead enrichment was entirely manual, creating bottlenecks in outreach, poor data fidelity, and high bounce rates.
solution:Developed multi-source lead ingestion and an API enrichment layer. Orchestrated CRM syncing, Slack notifications, lead scoring, and automated data validation.
Pythonn8nWeb ScrapingREST APIsCRM IntegrationsPostgreSQL
8h → 2h
enrichment / week
+30–50%
qualified leads
−40%
outreach delay
↑
CRM consistency
OPS-004Machine Learning Engineershipped
AsanaBot — Real-Time Yoga Pose Analysis

Domain-specific applied ML and vision: real-time yoga pose classification with corrective feedback.

challenge:Need for specialized, real-time guidance in physical yoga posture correction with instant feedback.
solution:Built AsanaBot using Vision Transformers and MediaPipe for accurate, real-time yoga pose classification and corrective feedback. Implemented efficient video processing and pose detection algorithms.
PythonVision TransformersMediaPipeOpenCVNLP
real-time
visual guidance
high
pose accuracy
instant
corrective signals
efficient
video processing
OPS-005Builder — automation · multi-agent · LLMinternal
Cold Outbound, Grounded in Research

For each lead, parallel Opus sub-agents pull what they can find on the person and company across the open web. A drafter writes an email that only uses hooks the agents actually cited — anything outside the evidence gets cut. Internal: drafts go to a review queue, not the SMTP.

challenge:Cold email is a tradeoff. Hand-researched emails take 10–20 minutes per lead. Templates scale but get ignored. And an LLM handed a name and a LinkedIn URL invents details — fake hooks that read worse than generic. I wanted the volume of templates without the made-up parts.
solution:Claude Code orchestrates the run. Each sub-agent has TypeScript tool scripts — Serper, LinkedIn / profile data, site scraping, LeadMagic, Prospeo, and a few public-record sources. Findings get consolidated into a research bundle per lead, then the drafter writes off that bundle only — no recalling from the model's own knowledge. Addresses are verified with MillionVerifier before anything is queued.
Claude CodeClaude OpusTypeScriptSerperLeadMagicProspeoMillionVerifierWeb Scraping
95.5%
faithful claims · 300-email sample
0
contradictions in sample
100%
agreement on 20-email calibration
evidence-bound
no hook without a cited fact
※Faithfulness is scored by a separate read-only eval that checks every claim in the draft against the research bundle. The harness caught a failure mode I'd missed — counting and naming things the evidence didn't actually support — which is now an explicit writing rule.
OPS-006Builder — scrapers · LLM classification · CI opsshipped
Tender Radar — UK Public-Sector Tender Pipeline

Autonomous pipeline that finds UK public-sector tenders competitors never see — and turned them into a paying-client funnel. Two OCDS API clients plus a Playwright scraper for portal-only notices, LLM sector classification behind a golden-set eval, and per-client morning digests.

challenge:UK facilities firms (cleaning, security, HVAC, grounds) miss winnable public contracts because school, council and NHS tenders are scattered across central government feeds AND regional eSourcing portals (ProContract and similar) that standard alert tools never read. A missed notice is a missed £100k–£7m contract.
solution:Built a fully unattended daily pipeline: OCDS fetchers for Find a Tender and Contracts Finder plus a Playwright scraper for portal-side tenders the APIs don't carry; LLM classification of every notice into client sectors behind a golden-set eval suite; a leads database with cross-source dedupe, word-boundary sector rules and deadline verification; per-client YAML profiles rendering a morning digest with deadline countdowns and portal-exclusive flags; GitHub Actions daily run plus a weekly 60-day national sweep, an award-notice tracker for contract-expiry intelligence, push-race guards and 12 pytest tests.
Python 3.11Pydantic v2PlaywrightGeminiOCDS APIspytestGitHub Actionsuv / ruff
7
sector verticals · all UK regions
78+
live verified tenders
100%
golden-set binary accuracy
0
manual steps · fetch → digest
※Production lessons: word-boundary sector matching, because "guard" inside "safeguarding" mistagged notices; an award-notice filter nearly killed a live £600k tender until the estimated-date case was fixed; keyword search endpoints silently ignore parameters for automated clients — only the OCDS bulk stream is trustworthy; CI push-races between the daily and weekly jobs lose data without rebase-retry guards.
// 04agents.swarm — what they look like running5 / 8 running
idagentmodeltaskstatuslat (s)
AGT-01lead-enrichergpt-4o-minienriching ACME Corp · 31 recordsrunning4.2
AGT-02resume-parserclaude-haikuranking 24 inbound CVsrunning2.1
AGT-03data-validatorpydantic-llmschema check · payroll batchidle—
AGT-04crm-syncern8n-runnerpushing 87 leads → hubspotrunning0.8
AGT-05slack-routerrule-enginerouting #gtm-warm-leadsqueued—
AGT-06doc-readerclaude-opusanalyzing Q1 founder updatesrunning11.3
AGT-07embed-indexerbge-largeindexing 1,240 chunksrunning0.4
AGT-08audit-loggerrule-engineno inbound eventsidle—
// 05stack.modules — what I build with28 modules · 6 groups

Tools that ship.

The libraries, languages, and platforms that show up across production codebases — not the ones I've only touched in a tutorial.
group.00
Core Languages
05
PythonCJavaScriptSQLTypeScript
group.01
Frameworks
04
Next.jsReactTailwind CSSFastAPI
group.02
AI & ML
06
LLM IntegrationPrompt EngineeringEmbeddingsVector SearchTransformer ModelsVision Transformers
group.03
Data & Libraries
05
PostgreSQLETL PipelinesPandasScikit-learnData Normalization
group.04
Automation & DB
04
n8nProcess AutomationLead EnrichmentCRM Automation
group.05
Tools
04
DockerBackground WorkersDeployment PipelinesGit
// 06execution.log — where this came from

Where this
came from.

Two recent engagements where I built and shipped the kind of system I’d build for you. Short timeline, real production code, real ops on the other side.

Precise Leads
GTM Automation Engineer
Nov 2025 – Present
Remote
missionIncrease GTM velocity by reducing human bottlenecks and operational overhead.
execution
  • Automating lead enrichment workflows and structuring CRM sync.
  • Integrating APIs for high-volume data processing.
  • Built a high-intent inbound discovery engine and scalable backend systems for immediate action.
impact
  • Reduced manual research time by 40%.
  • Delivered faster outreach cycles and higher data quality.
  • Improved overall GTM efficiency by replacing manual work with intelligent systems.
Caprae Capital Partners
AI / ML Intern
Jun 2025 – Dec 2025
Remote · Glendale, US
missionAutomate the candidate lifecycle and normalize high-volume financial data.
execution
  • Spearheaded the ground-up development of a proprietary recruitment ecosystem.
  • Engineered an AI-driven resume parsing and ranking system.
  • Architected candidate scoring models powered by advanced embeddings.
impact
  • Eliminated 100% of manual administrative friction in the hiring process.
  • Reduced screening time and drastically increased candidate throughput.
  • Improved candidate ranking consistency using systematic AI scoring logic.
edu: B.Tech, Computer Science · Maharaja Surajmal Institute of TechnologyNov 2022 – Aug 2026
// 07writing.log — notes from the build1 post

How I
actually build.

Engineering write-ups with the real numbers — failure modes, what fixed them, and the parts most demos skip.
01Jun 14, 2026·7 min read#llm#evals#pydantic

Schema-as-eval-spec: killing LLM failure modes by construction (and catching my eval cheating)

I built an LLM job-posting classifier, fixed three failure modes by changing the output schema instead of the prompt, then discovered my own evaluation was leaking the answer. The second part taught me more than the first.

cat ./schema-as-eval-spec.md →
// 08comms.open — start the engagement2 slots open
$ ./request_audit.sh --listen --no-fluff

Workflow eating hours?
Let’s kill it.

30-min audit, no slides, no fluff. Bring the workflow your team is sick of. I’ll sketch the architecture, scope the build, and quote the retainer to keep it running — in one call.

what i build / things i ship
S-01Lead enrichment & GTM pipelines
multi-source ingestion, scoring, CRM sync
S-02Document & resume parsing
NLP extraction, ranking, vector-DB search
S-03Multi-agent workflow systems
LLM agents that retry, escalate, audit
S-04Internal-tool automation
Slack · email · n8n · scheduled jobs
S-05Build + ongoing maintenance
ship it, then keep it running (retainer)
engagement
build + retainer
reply window
< 24h
timezone
IST · works async
next opening
2 slots this quarter