skip to content

dossier · no. 001 · luis osio chico · colombia, utc-5

Luis Osio Chico

ai engineer

for systems that answer with sources, or don't answer at all.

3+ years building production web platforms. Now the AI systems on top.

PythonFastAPITypeScriptReactNode.jsAWS

open to remote roles · utc-5 overlapapi · pinging…

Luis Osio Chico
remote-first
ask the fileit answers from the file — or says it can’t

answers only from this site's corpus

1/3
01 / work

Problems solved with AI

The assistant this site runs on, then two public builds on open corpora: a supplements-evidence RAG over PubMed research and a support-call → ticket pipeline over real call audio. Same rule as my production work: the machine drafts, evidence and a human decide. Repos go public with their eval reports.

128.6h diarized audio~92k transcript segmentsCPC H01M bounded corpus

ask the file

Live

retrieval workflow › the assistant this site runs on

The problem: Portfolio claims are easy to write and hard to check. A visitor shouldn't have to take my word for any of it.

The outcome: Ask it anything about my work. Every answer cites the file it came from, verified before it reaches you, or it abstains.

Five checked stages streamed live (gate, retrieve, scope, route, verify) over a corpus generated from this site's own machine layer. The writer answers under a contract of claims plus cited chunks. A judge scores every claim against its sources, and a failed check gets one repair pass or the run abstains.

PythonFastAPIWorkers AID1
One run, replayed

“what’s Luis’s stack?”

skills0

it routes, it doesn't grep — every other lane under 1%

tap or hover the bar — the gray sliver at the end is all 8 of them

the rest — all under 1%

about_me0.0090
projects0.0030
meta0.0009
contact0.0007
credentials0.0005
experience0.0005
restricted0.0002
off_topic0.0002

bars scaled to the largest runner-up (0.009)

wall clock · the writer call is 66% of it0

gate1/5189ms · 14%classifies the question — before retrieval→ 0.98 vs 0.009 — won by ~109×retrieve2/57ms · 1%fused top-k — 9 chunks, +3 canonical→ 12 chunks handed to scopescope3/5109ms · 8%re-judges evidence coverage→ pass · judge floor 0.18route4/5883ms · 66%the writer call — deepseek-v4.1-flash→ 66% of the wall clockverify5/5142ms · 11%every claim scored vs its source→ min claim 0.965 ≥ 0.70
gate 189msretrieve 7msscope 109msroute 883msverify 142ms

verify: 1 sentence, 1 sourcegrounded 0.965

✓

Luis's recorded stack is Python, FastAPI, TypeScript, React, Node.js, AWS, PostgreSQL, and MCP.

checked againstme.stackscore 0.965 ≥ 0.70

the answer's score is its weakest claim · under 0.70: one rewrite, then abstain

real numbers · one recorded runrun your own ↑
supplements, verifiedhow it works ↗

live — ask it anything; it only answers what the studies back

Supplements evidence RAG

In progress

retrieval workflow › studies, cited

The problem: Everyone asks 'does this supplement work?' — the honest answer lives in research papers nobody reads.

The outcome: A question in, the evidence back with PMID citations, a boundary when the corpus can't answer, a refusal when it's medical advice.

A RAG workflow in Python over a deliberately bounded corpus: dietary-supplement research on PubMed (~17.7k evidence-grade articles, 2015–2026). An advice gate declines medical advice; verification cites the PMID or abstains.

PythonWorkers AICloudflare Workers
1. Call transcript › speaker labelsin build

agent    "I can get a replacement shipped by Friday."

customer "it's the second time this part failed."

2. Pydantic schema contractvalid, 1 field → supervisor (confidence 0.58)

ticket #4821, every field cites its line

1 field in review
Issue
Repeat part failure, escalated
← customer, 04:12 "the second time this part failed"
Promised action
Replacement ships by Friday
← agent, 05:31 "I can get a replacement shipped by Friday."
Sentiment
Frustrated, escalation risk supervisor
← multi-turn signal, low confidence
Corpus: AppTek diarized audio, 128.6h → no quote, no fieldspec exhibit

Call → ticket

In progress

workflow with human review › call audio

The problem: Support calls hold every fact a ticket needs, trapped in audio nobody re-listens to.

The outcome: Audio in, a validated ticket out. Each field traceable to a verbatim line and timestamp; uncertain fields route to a supervisor.

Audio → ingest queue → speech-to-text with speaker labels → LLM field extraction → Pydantic gate → supervisor on uncertain fields → dispatched ticket. A promise must quote the agent's turn, asking isn't promising.

Speech-to-text · diarizationPydanticAWS

previously: in production at TXEN (private)

Péntica

In production (private)

retrieval workflow › tax rules, cited

Hybrid retrieval workflow over Colombian municipal tax statutes. Answers with the exact article cited, or abstains.

private client work, no public case file

Doru

In production (private)

workflow with human review › clinical records

Workflow that drafts a veterinary clinical record from consultation audio. A veterinarian reviews and approves every record.

private client work, no public case file

02 / tools

Tools AI assistants can use

MCP is the common thread: each of these exposes its function to agents, not just to people.

ex. 01

mdingest

Live

Turns articles and newsletters into clean Markdown for LLMs. One ingestion core behind an HTTP API, a CLI, and an MCP server. An article in, Markdown out: against the production API.

live on cloudflare · 4 entry points · mcp + llms.txt ready

TypeScriptBunCloudflare WorkersAstro
fig. 1 — system mapdeployedrequest in flight
GET /v1/*?url=http apimdingest <url>climcp · stdiolocal agentsmcp · httpremote agentsworker → doproxy · no logicingest(url)workers · one coredevtojson apisubstack/api/v1mediummirror
idle live request
—ingests served
ex. 01a — run it live
$
detected from url

> calls the production api at mdingest.knightker.workers.dev · real article in, markdown out. the url is yours to change.

ex. 02

italki-cli

Live

italki has no official API, so I built one from the outside: a Zod-validated CLI + MCP server to search, compare, and book lessons. The replay is a captured run: actual teachers, prices, and match counts.

13 mcp tools · 50+ fields zod-validated · arch enforced by lint

TypeScriptBunZodMCP

$ bun run index.ts mcp # stdio server · 13 tools

$ bun run verify # typecheck + 5 arch lint rules

italki-cli · recorded run, real api response0ms

$ _

Found 645 english teachers (type=pro, max=$20.00, native)   #32651986  Ethan M [PRO] from US · from $16.00 ★5 20 sessions trial $10.00  #4611201  Matt [PRO] from GB · from $15.00 new trial $8.00  #1084019  Lewella [PRO] from US · from $20.00 ★5 20178 sessions trial $5.00   Next page: italki search english --type pro --max-price 20 --native --page 2  View profile: italki teacher <id>

captured sep 28 · 645 real matches, real teachers, real prices

this site's MCP server

Live

This portfolio is itself callable: POST /v1/mcp. The full exchange runs in the machines section below.

03 / experience

The record

TXEN

AI Engineer

dec 2025 · present

~150ms

company-selector load, down from 5–8s · measured via CloudWatch

Software consultancy. AI Engineer on Péntica (B2B SaaS for municipal tax compliance across Colombia's 1,103 municipalities); previously on Doru (AI clinical-records SaaS for veterinary clinics).

  • Engineered Doru's AI consultation-to-clinical-record pipeline (speech-to-text, LLM structured extraction, Pydantic schema validation with automatic retries, and veterinarian human-in-the-loop review before records are saved); owned its event-driven automated reminders, AWS infrastructure, and React frontend optimizations
  • Built a hybrid RAG service in Python and FastAPI over Colombian tax regulations for accountant lookup (keyword + pgvector retrieval fused with RRF, article-aware chunking preserving each regulation's hierarchy, cross-encoder reranking, exact citations, superseded-norm detection, deterministic abstention on out-of-corpus or superseded norms), validated against a golden dataset curated by the accounting team
  • Cut company-selector load time from 5-8s to ~150ms across 14 frontend modules via a lightweight endpoint returning only required fields, measured via CloudWatch
  • Designed and executed a 19-phase PostgreSQL migration (Simplica to Péntica) across 1,103 municipalities with idempotent natural-key deduplication, dry-run mode, and per-municipality validation; built continuous SQS reverse sync and automated the ICA tax-matrix update email + file-attachment workflow
  • Migrated production to a new AWS account in under 12 hours (RDS snapshot + S3 sync), configured RDS Proxy + Lambda init-phase TypeORM loading to cut cold starts, and reduced 3,400 ESLint warnings to zero with Prettier team-wide
app
TypeScriptNode.jsPythonFastAPIReactPostgreSQLTypeORM
aws
AWS LambdaAPI GatewayRDSRDS ProxyS3SQSCloudWatch
ship
SST v3GitHub Actions

Nord Tech

Software Engineer

dec 2024 · dec 2025

<12mo

to promotion from Junior

Promoted from Junior inside a year on a real-time reservation and streaming platform.

  • Delivered backend work on a real-time reservation and streaming platform in Scala + Kafka, under a senior lead's direction.
  • Reviewed code and assigned tasks for two junior developers, supporting their delivery.
app
TypeScriptNode.jsPythonReactScalaKafkaWebSockets
aws
AWS Amplify
↑ promoted

Junior Software Engineer

jan 2024 · dec 2024

100+

endpoints migrated, zero incidents

First role. Migrated a lending platform's backend off end-of-life Node 14.

  • Migrated 100+ endpoints (Node.js/TypeScript on AWS Lambda): built the unit + integration test foundation around business-critical endpoints, deployed with zero incidents
  • Earned the promotion on code quality, autonomy, and consistent delivery
app
TypeScriptNode.jsPythonReact
aws
AWS AmplifyAWS Lambda

No Country

Frontend Developer (Intern)

jun 2023 · sep 2023

2

project builds shipped during the internship

Project-based internship: real team deliverables on two LATAM community builds.

  • Urban Move (urban mobility): responsive UIs in React, TypeScript, Tailwind CSS, Vite; frontend testing with Vitest
  • UVA (developer-project matching): frontend in React, Vite; code quality via ESLint + Prettier
app
ReactTypeScriptTailwind CSSViteVitest
04 / references

What people I worked with say

LinkedIn recommendations, translated from Spanish. originals on file

“Luis showed strong technical judgment building with AI-assisted tools. He's always at the leading edge. His professional ethics and his ability to communicate across teams and with clients stand out.”
Jose Daniel Posso Garcia

Jose Daniel Posso Garcia

CTO & Founder, Doru, TXEN · managed luis directly

“One of the best developers I know. Before implementing anything he evaluates the options, understands the context, and makes informed decisions. I'd recommend him to any team, remote included.”
Jhovani Moreno

Jhovani Moreno

Full Stack Developer · same team, pentica

“High technical ability and excellent problem-solving skills, with real commitment to the quality of the work. Responsible, proactive, with strong professional ethics.”
Michel Andrey Gallo Perez

Michel Andrey Gallo Perez

Software Engineer · same team, pentica

“Across Next.js, AWS Lambda, TanStack and DynamoDB he showed real capacity for technical challenges, and proposed solutions oriented to what the project actually needed.”
Mateus Tegue

Mateus Tegue

Full Stack Developer · different team, doru

05 / credentials

Education & certifications

Software Engineering (BSc)

UniHorizonte, Colombia

in progress, 2022, exp. 2027

Oracle Next Education, Back-End

Alura LATAM, 286h, 6 courses

earned, 2024

Web Development & Programming

Platzi

earned

06 / machines

This site can also be read by AI assistants

Every page also exists as structured data, so agents and recruiters read the same truth: /v1/me.json, llms.txt, openapi.json. The card on the right runs a real request from your browser.

And it's callable, not just readable: POST /v1/mcp: this dossier is itself an MCP server. The config drops into claude, cursor, or any mcp client.

luisosio.dev · mcp server

→ POST /v1/mcp

{"jsonrpc":"2.0","id":1,"method":"tools/list"}

← 4 tools · test it live for a real answer

  • get_profileprofile, stack, links
  • list_projectsshipped work, repos, case-study links
  • get_experienceroles, companies, production metrics
  • get_projectone project's full case study, markdown
verified at build
luisosio.dev · response body

$ curl luisosio.dev/v1/me.json

→ GET /v1/me.json
{
  "name": "Luis Osio Chico",
  "handle": "losiochico",
  "role": "AI Engineer",
  "stack": [
    "Python",
    "FastAPI",
    "TypeScript",
    "React",
    "Node.js",
    "AWS",
    "PostgreSQL",
    "MCP"
  ],
  "focus": "AI systems that answer with sources or abstain: retrieval workflows, MCP tools, human-in-the-loop validation",
  "location": "Colombia",
  "timezone": "America/Bogota (UTC-5)",
  "status": "open_to_work",
  "links": {
    "github": "https://github.com/LOsioChico",
    "linkedin": "https://linkedin.com/in/losiochico",
    "email": "[email protected]",
    "cv": "https://luisosio.dev/cv.pdf"
  },
  "endpoints": [
    "/v1/me.json",
    "/v1/projects.json",
    "/v1/experience.json",
    "/llms.txt",
    "/.well-known/ai-catalog.json"
  ]
}

↳ run a random one → real call

07 / contact

Need AI that says
"I don't know" when it should?

Open to AI engineer roles: remote, LATAM and US time zones. Retrieval workflows, MCP tools, and human-in-the-loop validation on a production spine. the full hire page →

[email protected]
github linkedindownload cv ↗