Hi
Hi
Allan Dsouza
A I
Solutions
Engineer
I build agentic systems — the workflows, memory layers, and guardrails that let AI run real business processes.
I build agentic systems — the workflows, memory layers, and guardrails that let AI run real business processes.
what I can do for you
I design and run agentic systems end to end — multi-step workflows with human approval gates, retrieval and memory layers agents can actually rely on, and the evals, budgets and fallbacks that keep them reliable once they are live.
1. Agent Orchestration
Stateful multi-step workflows and tool calling
Human-in-the-loop approval gates
Webhook and schedule triggers
Bounded loops with budgets and graceful failure
2. Retrieval & Memory
Chunking, embeddings and vector search
Reranking and hybrid retrieval
Knowledge graphs and organisational memory
Citation verification and anti-hallucination
3. AI Development
LLM integration (OpenAI, Groq, Claude)
RAG systems and vector databases
AI agents and MCP servers
Prompt engineering and fine-tuning
4. AI Automation
Workflow automation with AI agents
GitHub issue triage and classification
Content generation pipelines
Custom AI tools for productivity
1. Agent Orchestration
Stateful multi-step workflows and tool calling
Human-in-the-loop approval gates
Webhook and schedule triggers
Bounded loops with budgets and graceful failure
2. Retrieval & Memory
Chunking, embeddings and vector search
Reranking and hybrid retrieval
Knowledge graphs and organisational memory
Citation verification and anti-hallucination
3. AI Development
LLM integration (OpenAI, Groq, Claude)
RAG systems and vector databases
AI agents and MCP servers
Prompt engineering and fine-tuning
4. AI Automation
Workflow automation with AI agents
GitHub issue triage and classification
Content generation pipelines
Custom AI tools for productivity
About me
Hi, I'm Allan. At Edmik I own the AI layer of a production platform — model routing with fallback, token budgets, and an LLM grading pipeline. Alongside it I ship agentic systems. Five years designing games taught me that failure handling is the product, which is exactly how I think about agents.
8+
8+
Years in Game Design
8+
8+
Years in Game Design
8+
8+
AI Projects Shipped
8+
8+
AI Projects Shipped
8+
8+
Open Source Projects
8+
8+
Open Source Projects
Email :
allands365@gmail.com
Email :
allands365@gmail.com
Featured Projects
Agents wired into real systems, with explicit blast radius and numbers behind the reliability claims. Each of these runs, and each one has a public repo.
Agentic System
Distill — AI Triage & Release Operator
A multi-step agent workflow that turns raw feedback from email, Slack and GitHub into classified, prioritised, deduplicated issues with reproduction steps — routing low-confidence and P0 items to a human approval gate. Priority scoring and dedup are deterministic code, not prompts. Wired into Gmail, Slack and Linear, with webhook-triggered auto-triage.
Knowledge & Memory
MindMeld — Cross-Meeting Knowledge Graph
Records and transcribes meetings, then folds each one into a single knowledge graph spanning every meeting — using the full memory lifecycle: ingest, recall across meetings, self-organising enrichment, and forget. Ask whether a decision made today contradicts one made at kickoff, and it reasons across both to answer.
Agent Reliability
KSP Crime Intelligence — Bounded Agent Loop
A natural-language analyst agent over 674k records, in English and Kannada. Its tool schemas are derived from function signatures so they cannot drift from the code, and a population test loops the full registry so a new tool cannot skip coverage. Capped at two rounds with a shared latency budget. Measured, not asserted: 8/10 correct tool selection, 10/10 within budget.
Applied ML
MSME Financial Health Card
Explainable credit scoring for credit-invisible small businesses from alternate data. A WOE/logistic scorecard trained offline and applied in TypeScript from a committed artifact, with an XGBoost + SHAP challenger split out to an ONNX microservice so the deployment target stayed an open decision. 55 unit tests, synthetic-by-design data.
Frequently Asked Questions
How I think about building agents — where they earn autonomy, where they shouldn't exist at all, and how I know they still work once they're live. If your question isn't here, reach out.
1.
What do you actually build?
2.
What's your stack?
3.
When should something be an agent — and when shouldn't it?
4.
How do you stop an agent doing damage?
5.
How do you know an agent is actually reliable?
6.
What are you looking for?
1.
What do you actually build?
2.
What's your stack?
3.
When should something be an agent — and when shouldn't it?
4.
How do you stop an agent doing damage?
5.
How do you know an agent is actually reliable?
6.
What are you looking for?
1.
What do you actually build?
2.
What's your stack?
3.
When should something be an agent — and when shouldn't it?
4.
How do you stop an agent doing damage?
5.
How do you know an agent is actually reliable?
6.
What are you looking for?
Hi
Hi
Let's work together
I'm open to AI engineering roles — agentic systems, retrieval and memory, and the infrastructure that keeps them reliable. Based in Mumbai, comfortable with US-overlapping hours.
Hi
Hi
Allan Dsouza
A I
Solutions
Engineer
I build agentic systems — the workflows, memory layers, and guardrails that let AI run real business processes.
I build agentic systems — the workflows, memory layers, and guardrails that let AI run real business processes.
what I can do for you
I design and run agentic systems end to end — multi-step workflows with human approval gates, retrieval and memory layers agents can actually rely on, and the evals, budgets and fallbacks that keep them reliable once they are live.
1. Agent Orchestration
Stateful multi-step workflows and tool calling
Human-in-the-loop approval gates
Webhook and schedule triggers
Bounded loops with budgets and graceful failure
2. Retrieval & Memory
Chunking, embeddings and vector search
Reranking and hybrid retrieval
Knowledge graphs and organisational memory
Citation verification and anti-hallucination
3. AI Development
LLM integration (OpenAI, Groq, Claude)
RAG systems and vector databases
AI agents and MCP servers
Prompt engineering and fine-tuning
4. AI Automation
Workflow automation with AI agents
GitHub issue triage and classification
Content generation pipelines
Custom AI tools for productivity
1. Agent Orchestration
Stateful multi-step workflows and tool calling
Human-in-the-loop approval gates
Webhook and schedule triggers
Bounded loops with budgets and graceful failure
2. Retrieval & Memory
Chunking, embeddings and vector search
Reranking and hybrid retrieval
Knowledge graphs and organisational memory
Citation verification and anti-hallucination
3. AI Development
LLM integration (OpenAI, Groq, Claude)
RAG systems and vector databases
AI agents and MCP servers
Prompt engineering and fine-tuning
4. AI Automation
Workflow automation with AI agents
GitHub issue triage and classification
Content generation pipelines
Custom AI tools for productivity
About me
Hi, I'm Allan. At Edmik I own the AI layer of a production platform — model routing with fallback, token budgets, and an LLM grading pipeline. Alongside it I ship agentic systems. Five years designing games taught me that failure handling is the product, which is exactly how I think about agents.
8+
8+
Years in Game Design
8+
8+
Years in Game Design
8+
8+
AI Projects Shipped
8+
8+
AI Projects Shipped
8+
8+
Open Source Projects
8+
8+
Open Source Projects
Email :
allands365@gmail.com
Email :
allands365@gmail.com
Featured Projects
Agents wired into real systems, with explicit blast radius and numbers behind the reliability claims. Each of these runs, and each one has a public repo.
Agentic System
Distill — AI Triage & Release Operator
A multi-step agent workflow that turns raw feedback from email, Slack and GitHub into classified, prioritised, deduplicated issues with reproduction steps — routing low-confidence and P0 items to a human approval gate. Priority scoring and dedup are deterministic code, not prompts. Wired into Gmail, Slack and Linear, with webhook-triggered auto-triage.
Knowledge & Memory
MindMeld — Cross-Meeting Knowledge Graph
Records and transcribes meetings, then folds each one into a single knowledge graph spanning every meeting — using the full memory lifecycle: ingest, recall across meetings, self-organising enrichment, and forget. Ask whether a decision made today contradicts one made at kickoff, and it reasons across both to answer.
Agent Reliability
KSP Crime Intelligence — Bounded Agent Loop
A natural-language analyst agent over 674k records, in English and Kannada. Its tool schemas are derived from function signatures so they cannot drift from the code, and a population test loops the full registry so a new tool cannot skip coverage. Capped at two rounds with a shared latency budget. Measured, not asserted: 8/10 correct tool selection, 10/10 within budget.
Applied ML
MSME Financial Health Card
Explainable credit scoring for credit-invisible small businesses from alternate data. A WOE/logistic scorecard trained offline and applied in TypeScript from a committed artifact, with an XGBoost + SHAP challenger split out to an ONNX microservice so the deployment target stayed an open decision. 55 unit tests, synthetic-by-design data.
Frequently Asked Questions
How I think about building agents — where they earn autonomy, where they shouldn't exist at all, and how I know they still work once they're live. If your question isn't here, reach out.
1.
What do you actually build?
2.
What's your stack?
3.
When should something be an agent — and when shouldn't it?
4.
How do you stop an agent doing damage?
5.
How do you know an agent is actually reliable?
6.
What are you looking for?
1.
What do you actually build?
2.
What's your stack?
3.
When should something be an agent — and when shouldn't it?
4.
How do you stop an agent doing damage?
5.
How do you know an agent is actually reliable?
6.
What are you looking for?
1.
What do you actually build?
2.
What's your stack?
3.
When should something be an agent — and when shouldn't it?
4.
How do you stop an agent doing damage?
5.
How do you know an agent is actually reliable?
6.
What are you looking for?
Hi
Hi
Let's work together
I'm open to AI engineering roles — agentic systems, retrieval and memory, and the infrastructure that keeps them reliable. Based in Mumbai, comfortable with US-overlapping hours.
Hi
Hi
Allan Dsouza
A I
Solutions
Engineer
I build agentic systems — the workflows, memory layers, and guardrails that let AI run real business processes.
I build agentic systems — the workflows, memory layers, and guardrails that let AI run real business processes.
what I can do for you
I design and run agentic systems end to end — multi-step workflows with human approval gates, retrieval and memory layers agents can actually rely on, and the evals, budgets and fallbacks that keep them reliable once they are live.
1. Agent Orchestration
Stateful multi-step workflows and tool calling
Human-in-the-loop approval gates
Webhook and schedule triggers
Bounded loops with budgets and graceful failure
2. Retrieval & Memory
Chunking, embeddings and vector search
Reranking and hybrid retrieval
Knowledge graphs and organisational memory
Citation verification and anti-hallucination
3. AI Development
LLM integration (OpenAI, Groq, Claude)
RAG systems and vector databases
AI agents and MCP servers
Prompt engineering and fine-tuning
4. AI Automation
Workflow automation with AI agents
GitHub issue triage and classification
Content generation pipelines
Custom AI tools for productivity
1. Agent Orchestration
Stateful multi-step workflows and tool calling
Human-in-the-loop approval gates
Webhook and schedule triggers
Bounded loops with budgets and graceful failure
2. Retrieval & Memory
Chunking, embeddings and vector search
Reranking and hybrid retrieval
Knowledge graphs and organisational memory
Citation verification and anti-hallucination
3. AI Development
LLM integration (OpenAI, Groq, Claude)
RAG systems and vector databases
AI agents and MCP servers
Prompt engineering and fine-tuning
4. AI Automation
Workflow automation with AI agents
GitHub issue triage and classification
Content generation pipelines
Custom AI tools for productivity
About me
Hi, I'm Allan. At Edmik I own the AI layer of a production platform — model routing with fallback, token budgets, and an LLM grading pipeline. Alongside it I ship agentic systems. Five years designing games taught me that failure handling is the product, which is exactly how I think about agents.
8+
8+
Years in Game Design
8+
8+
Years in Game Design
8+
8+
AI Projects Shipped
8+
8+
AI Projects Shipped
8+
8+
Open Source Projects
8+
8+
Open Source Projects
Email :
allands365@gmail.com
Email :
allands365@gmail.com
Featured Projects
Agents wired into real systems, with explicit blast radius and numbers behind the reliability claims. Each of these runs, and each one has a public repo.
Agentic System
Distill — AI Triage & Release Operator
A multi-step agent workflow that turns raw feedback from email, Slack and GitHub into classified, prioritised, deduplicated issues with reproduction steps — routing low-confidence and P0 items to a human approval gate. Priority scoring and dedup are deterministic code, not prompts. Wired into Gmail, Slack and Linear, with webhook-triggered auto-triage.
Knowledge & Memory
MindMeld — Cross-Meeting Knowledge Graph
Records and transcribes meetings, then folds each one into a single knowledge graph spanning every meeting — using the full memory lifecycle: ingest, recall across meetings, self-organising enrichment, and forget. Ask whether a decision made today contradicts one made at kickoff, and it reasons across both to answer.
Agent Reliability
KSP Crime Intelligence — Bounded Agent Loop
A natural-language analyst agent over 674k records, in English and Kannada. Its tool schemas are derived from function signatures so they cannot drift from the code, and a population test loops the full registry so a new tool cannot skip coverage. Capped at two rounds with a shared latency budget. Measured, not asserted: 8/10 correct tool selection, 10/10 within budget.
Applied ML
MSME Financial Health Card
Explainable credit scoring for credit-invisible small businesses from alternate data. A WOE/logistic scorecard trained offline and applied in TypeScript from a committed artifact, with an XGBoost + SHAP challenger split out to an ONNX microservice so the deployment target stayed an open decision. 55 unit tests, synthetic-by-design data.
Frequently Asked Questions
How I think about building agents — where they earn autonomy, where they shouldn't exist at all, and how I know they still work once they're live. If your question isn't here, reach out.
1.
What do you actually build?
2.
What's your stack?
3.
When should something be an agent — and when shouldn't it?
4.
How do you stop an agent doing damage?
5.
How do you know an agent is actually reliable?
6.
What are you looking for?
1.
What do you actually build?
2.
What's your stack?
3.
When should something be an agent — and when shouldn't it?
4.
How do you stop an agent doing damage?
5.
How do you know an agent is actually reliable?
6.
What are you looking for?
1.
What do you actually build?
2.
What's your stack?
3.
When should something be an agent — and when shouldn't it?
4.
How do you stop an agent doing damage?
5.
How do you know an agent is actually reliable?
6.
What are you looking for?
Hi
Hi
Let's work together
I'm open to AI engineering roles — agentic systems, retrieval and memory, and the infrastructure that keeps them reliable. Based in Mumbai, comfortable with US-overlapping hours.