Pengulab
Building experimental reasoning systems
We build MIR-NN, an experimental neuro-symbolic AI architecture designed to reason over structured problems, test hypotheses, analyze errors, use auditable memory, and improve through verification-driven reasoning loops.
Experimental AI systems for reasoning, data analysis, verification, and adaptive problem solving.
What is Pengulab?
Pengulab is an AI lab focused on building experimental reasoning architectures. Our goal is to explore AI systems that do more than generate text: systems that can represent problems, decompose them, propose hypotheses, verify results, analyze residual errors, store useful traces, and repair their own reasoning process.
MIR-NN: Reasoning Beyond Text Prediction
MIR-NN is our core experimental architecture. It combines neural networks with structured reasoning, explicit verifiers, residual analysis, auditable memory, decomposition views, and internal cognitive actions.
Instead of following a simple input → output pipeline, MIR-NN follows a reasoning loop:
Universal State Representation
Transforms different problems into a common structured reasoning format.
ThoughtToken Actions
The system can generate internal cognitive actions such as decomposing, comparing, verifying, repairing, and searching.
Explicit Verification
Hypotheses are tested instead of simply accepted because they sound plausible.
Residual Analysis
Errors are analyzed as structured signals that guide the next attempt.
Auditable Memory
The system stores useful traces, failures, evidence, concepts, and counterexamples.
Multidomain Design
MIR-NN is designed to reason over data, tables, code, planning tasks, causal problems, language evidence, and abstract visual problems.
Data Mining
MIR-NN is an auditable AI reasoning engine designed to transform raw business data into structured, evidence-backed insights.
Most analytics tools treat data as something to visualize. Most AI chatbots treat data as something to summarize. MIR-NN approaches data differently: it treats every dataset as a reasoning problem that must be decomposed, analyzed, verified, and explained with evidence.
When MIR-NN receives a dataset, it transforms the data into a structured internal representation called UniversalState. Columns, rows, categories, dates, prices, quantities, regions, products, customers, and derived metrics become part of a reasoning state composed of entities, relationships, goals, actions, evidence, and verification steps.
A user may ask:
"Why did sales drop this month?"
Instead of generating a quick explanation, MIR-NN is designed to decompose the question:
Traditional AI Output
"Sales dropped because of Product A."
MIR-NN Output
- Revenue decreased by 18.4% month over month.
- The largest negative contribution came from Product A in the North Region.
- Product A explains 63% of the total revenue drop.
- Verification passed using columns: date, product, region, price, and quantity.
Anomaly Detection
MIR-NN can detect anomalies in structured datasets: unusual values, missing fields, duplicated entries, impossible dates, negative quantities, zero prices, sudden metric shifts, or category-level irregularities.
Anomaly detected:
Row 184 has revenue 4.8x higher than the category median.
Evidence: Category: hardware | Category median revenue: 120 | Row revenue: 576
Verification: Calculation passed.
Pattern Discovery
MIR-NN can identify products that rise or fall together, regions with similar behavior, customer groups with abnormal changes, correlated columns, segments with higher risk, temporal shifts, and groups that explain a metric change.
Pattern found:
Customers with plan = "basic" and usage above 80% show higher churn risk.
Evidence: Churn rate for basic high-usage customers: 34% | Overall churn rate: 12% | Lift: 2.83x
Verification: Segment calculation passed.
Data Analysis + Knowledge Retrieval
MIR-NN combines data analysis with strict retrieval over internal knowledge. It can analyze the dataset, verify findings, retrieve relevant internal notes or reports, cite evidence, and separate verified facts from possible explanations.
"Sales dropped in Mendoza. Is there any internal report explaining why?"
Verified data finding: Revenue in Mendoza decreased by 21% compared to the previous period.
Retrieved internal evidence: An internal operations note mentions a stock shortage in that region during the same period.
Conclusion: The stock shortage is a plausible explanation, but not causally proven from the available data.
Auditable Reports
MIR-NN generates reports where each conclusion is connected to the metric used, the calculation performed, the rows or columns involved, the evidence retrieved, the verification result, the confidence level, and the limitations of the analysis.
Conclusion: Revenue decreased by 18.4%.
Evidence: Previous period revenue: $124,000 | Current period revenue: $101,200 | Difference: -$22,800 | Variation: -18.4%
Verification: Calculation repeated successfully. Columns used: date, price, quantity.
Remaining uncertainty: The dataset confirms the revenue drop, but does not fully explain the cause. Recommended next analysis: stock levels, discounts, marketing traffic, and customer churn.
Different from a Dashboard
MIR-NN does not only display metrics. It asks which metric matters, searches for possible explanations, verifies calculations, stores reasoning traces, and separates evidence from assumptions.
Different from a Chatbot
MIR-NN does not simply generate a plausible explanation. It is designed to reason through the data, verify what it can, refuse unsupported claims, and recommend the next analysis when uncertainty remains.
MIR Language Model
MIR-NN is not another generic LLM.
Most language models are trained to predict the next word. MIR-NN is designed to do something different: transform language into structured reasoning.
It reads a user request, converts it into a structured internal state, generates cognitive action tokens, retrieves verified knowledge from memory or SQL databases, checks its own outputs, and only then produces a grounded answer.
The result is an AI system built for reliability, traceability, and decision support — not just fluent conversation.
What MIR-NN Is
MIR-NN is a verifier-guided reasoning model that combines:
Instead of relying only on hidden model weights, MIR-NN separates:
- - What the user asked
- - What the system understood
- - What evidence was retrieved
- - What reasoning actions were taken
- - What was verified
- - What remains uncertain
This makes MIR-NN more transparent than a standard black-box chatbot.
MIR-NN turns language generation into a verified reasoning process.
It is designed for teams that need AI systems that can answer questions, analyze information, retrieve evidence, explain decisions, and show what was actually verified.
Traditional LLMs
prompt → model weights → answerMIR-NN
prompt → language tokens→ UniversalState → ThoughtTokens→ SQL / memory / tools→ verification → grounded answerThe model weights are not used as a giant memory of everything.
They are used to understand language, select reasoning actions, connect concepts, and guide the system through a verifiable reasoning process.
The knowledge lives in structured memory and databases. The reasoning is traced. The answer is checked.
Why It Is Different
Standard AI assistants generate answers.
MIR-NN builds a reasoning trace.
Standard LLMs hide most knowledge in weights.
MIR-NN connects language to explicit SQL/database memory.
Standard chatbots may hallucinate.
MIR-NN is designed to verify, cite, or refuse.
Standard models respond directly.
MIR-NN structures, reasons, checks, and then responds.
Built for environments where hallucination is expensive
Why We Do This
We are passionate about artificial intelligence. We believe the path to AGI requires more than scaling: it requires systems that can reason, verify, explain, and learn from their own errors.
Pengulab exists because we want to build AI that thinks — not just AI that sounds like it thinks. Verifiable reasoning is the foundation of trustworthy intelligence.
Get in Touch
Interested in MIR-NN, partnerships, or just want to talk about reasoning AI?
Connect with us on LinkedIn
Pengulab on LinkedInor email us at pengulab@gmail.com